Re: CAN messages being lost on i.MX25 with flexcan - continued (was CAN messages being lost on i.MX25 with flexcan - 2012-04-19)

[Wolfgang Grandegger <wg@grandegger.com>]

On 10/29/2013 11:46 AM, Martin Kozusky wrote:
> Dne 26.10.2013 21:40, Wolfgang Grandegger napsal(a):
>> On 10/26/2013 08:18 PM, Martin Kozusky wrote:
>>> Dne 25.10.2013 19:58, Wolfgang Grandegger napsal(a):
>>>> Hi Martin,
>>>>
>>>> On 10/25/2013 02:59 PM, Martin Kozusky wrote:
>>>>> Dne 24.10.2013 15:48, Martin Kozusky napsal(a):
>>>>>> Hello,
>>>>>> after more than year I'm back with CAN fifo overrun problems on
>>>>>> i.MX25
>>>>>> board.
>>>>>> (it was good enough earlier, if some frames were lost, but not this
>>>>>> time)
>>>>>>
>>>>>> I have 2 flexcan interfaces, each receiving around 1100 msgs/s
>>>>>> (situation is a little better if I use just one iface, but I need
>>>>>> both)
>>>>>> I just configure them and then run:
>>>>>>
>>>>>> I=0; while [ $I -le 20 ]; do dd if=/dev/zero of=/mnt/mmcblk0p1/test
>>>>>> bs=512 count=200; sync; sleep 1; I=$(($I+1)); done
>>>>>> (simulate writing to SDcard with 100KB blocks in 1 sec intervals)
>>>>>>
>>>>>> and start sending data from another device.
>>>>>>
>>>>>> I am not running any other program (like candump etc) to read from
>>>>>> CAN.
>>>>>>
>>>>>> this is what is shown after I finish sending 35777 packets (both
>>>>>> interfaces now connected to same bus so they should receive same
>>>>>> data)
>>>>>> with ip -d -s link show can0/1
>>>>>>
>>>>>> 2: can0: <NOARP,ECHO> mtu 16 qdisc pfifo_fast state DOWN qlen 10
>>>>>>        link/can
>>>>>>        can <LISTEN-ONLY> state STOPPED (berr-counter tx 0 rx 0)
>>>>>> restart-ms 0
>>>>>>        bitrate 250000 sample-point 0.857
>>>>>>        tq 285 prop-seg 5 phase-seg1 6 phase-seg2 2 sjw 1
>>>>>>        flexcan: tseg1 4..16 tseg2 2..8 sjw 1..4 brp 1..256 brp-inc 1
>>>>>>        clock 66500000
>>>>>>        re-started bus-errors arbit-lost error-warn error-pass bus-off
>>>>>>        0          0          0          1          1          0
>>>>
>>>> Do you have electrical problems on the bus? Or is reaching
>>>> error-passive
>>>> not related to this problem?
>>> It is not related to this problem - there is only RX pin connected on
>>> can0 (RX is connected in parallel with Coldfire V1 MCU CAN, which is
>>> doing TX)
>>>
>>>
>>>>>>        RX: bytes  packets  errors  dropped overrun mcast
>>>>>>        151769     19000    1699    0       1699    0
>>>>>>        TX: bytes  packets  errors  dropped carrier collsns
>>>>>>        0          0        0       0       0       0
>>>>>> root@vmx25 /opt/waytracer$ /root/utils/ip -d -s link show can1
>>>>>> 3: can1: <NOARP,ECHO> mtu 16 qdisc pfifo_fast state DOWN qlen 10
>>>>>>        link/can
>>>>>>        can state STOPPED (berr-counter tx 0 rx 0) restart-ms 0
>>>>>>        bitrate 250000 sample-point 0.857
>>>>>>        tq 285 prop-seg 5 phase-seg1 6 phase-seg2 2 sjw 1
>>>>>>        flexcan: tseg1 4..16 tseg2 2..8 sjw 1..4 brp 1..256 brp-inc 1
>>>>>>        clock 66500000
>>>>>>        re-started bus-errors arbit-lost error-warn error-pass bus-off
>>>>>>        0          0          0          0          0          0
>>>>>>        RX: bytes  packets  errors  dropped overrun mcast
>>>>>>        157377     19696    2664    0       2664    0
>>>>>>        TX: bytes  packets  errors  dropped carrier collsns
>>>>>>        0          0        0       0       0       0
>>>>>>
>>>>>>
>>>>>> With just one iface used:
>>>>>>
>>>>>> 2: can0: <NOARP,ECHO> mtu 16 qdisc pfifo_fast state DOWN qlen 10
>>>>>>        link/can
>>>>>>        can <LISTEN-ONLY> state STOPPED (berr-counter tx 0 rx 0)
>>>>>> restart-ms 0
>>>>>>        bitrate 250000 sample-point 0.857
>>>>>>        tq 285 prop-seg 5 phase-seg1 6 phase-seg2 2 sjw 1
>>>>>>        flexcan: tseg1 4..16 tseg2 2..8 sjw 1..4 brp 1..256 brp-inc 1
>>>>>>        clock 66500000
>>>>>>        re-started bus-errors arbit-lost error-warn error-pass bus-off
>>>>>>        0          0          0          1          1          0
>>>>>>        RX: bytes  packets  errors  dropped overrun mcast
>>>>>>        233277     29201    1483    0       1483    0
>>>>>>        TX: bytes  packets  errors  dropped carrier collsns
>>>>>>        0          0        0       0       0       0
>>>>>>
>>>>>>
>>>>>> Too many packets are lost.
>>>>>>
>>>>>> I tried to play with FLEXCAN_NAPI_WEIGHT (quota for napi) and that
>>>>>> didn't hepl too much, if I put it too high then the system response
>>>>>> was slow and packets still lost, also tried to change priority of CAN
>>>>>> interrupts with (don't know if correctly)
>>>>>>      // imx_irq_set_priority(43,14);
>>>>>>      // imx_irq_set_priority(44,14);
>>>>>>
>>>>>> But it didn't help either.
>>>>>>
>>>>>>
>>>>>> Does anybody have any idea how not to lose any packets? :)
>>>>>
>>>>>
>>>>> Hello,
>>>>> I tried to disable
>>>>> //netif_receive_skb(skb); in  flexcan_read_frame() and other functions
>>>>> so that data is not processed further in system
>>>>
>>>> Well ...
>>>>
>>>>> It didn't help.
>>>>> So I tried to put time_start=ktime_get_real() at the begining of
>>>>> flexcan_read_frame(), then time_stop=ktime_get_real(); at the end and
>>>>> add their difference to the global variable
>>>>> time_total+=time_stop-time_start;
>>>>> I divided this time_total by rx_packets count at flexcan_chip_stop and
>>>>> wrote with dev_info into log (variables were initialized in
>>>>> flexcan_chi_start, so I could just do ifconfig can0 up/down and reset
>>>>> those counters and write them to log), so now I had average time spent
>>>>> int flexcan_read_frame.
>>>>> This time it was around 100usec! just with one CAN used, if both were
>>>>> connected, it was more than twice. And many CAN frames were lost.
>>>>>
>>>>> So I tried to disable
>>>>>       /*
>>>>>           skb = alloc_can_skb(dev, &cf);
>>>>>           if (unlikely(!skb)) {
>>>>>                   stats->rx_dropped++;
>>>>>                   return 0;
>>>>>           }
>>>>>        */
>>>>> and made "struct can_frame cf" (not pointer, so that I can use it in
>>>>> flexcan_read_fifo call)
>>>>> And tried to send data again.
>>>>> Now - average time in flexcan_read_frame was not 100usec, but just 2
>>>>> usec! 50x less ...  no CAN frame was lost, even if I was using both
>>>>> CAN
>>>>> interfaces, each getting over 1100 msgs/sec and writing 100KB data
>>>>> to SD
>>>>> card.
>>>>
>>>> ... but the messages need to be allocated, queued, delivered to and
>>>> even
>>>> processed by a user space task. What you messure it part of the network
>>>> stack overhead but 100us just for alloc_can_skb() seems quite a lot to
>>>> me. At what frequency is your CPU running? Is the system low of memory?
>>>> Maybe your system is simply not fast enough. To see what code is
>>>> involved just follow:
>>> CPU is i.MX25, should be running at 400MHz. There is 64MB RAM totally
>>> and free enough :(
>>>
>>>>
>>>>     http://lxr.free-electrons.com/ident?i=alloc_can_skb
>>>>> So I am asking - how to make this alloc_can_skb faster (or is there
>>>>> any
>>>>> alternative)? Or if there is another way how to get data to user?
>>>>
>>>> Well, not with Linux-CAN. Anyway, messages arrive at a rate of
>>>> approx. 1
>>>> KHz. So there is 1ms per message. I think it's a latency problem in the
>>>> first place. The Flexcan on the i.MX25 can queue up to 5 messages. If
>>>> the queue is full you loose messages. This obviously happens when the
>>>> SDcard is accessed.
>>>>
>>>> Could you take function traces on your system?
>>> Is there any special tool for this or should I use my start/stop timers?
>>
>> Your start/stop timers will not show what other activity is disturbing
>> the CAN messages reception. There is the Linux function tracer:
>>
>>    http://lxr.free-electrons.com/source/Documentation/trace/ftrace.txt
>>
>> It need to be enabled in the kernel. Especially event and function
>> tracing could help to better understand your problems.
> Hello Wolfgang,
> it seems that my architecture (arm/mx25 on 2.6.35 kernel) is missing
> HAVE_FUNCTION_GRAPH_TRACER, HAVE_DYNAMIC_FTRACE options so it won't be
> that easy, will be?
> Timestamps that ftrace is showing me are in 10 miliseconds resolution,
> that won't help me much :(

Probably that version is to old for proper ftrace support. The 100us you
measured for alloc_can_skb() is worst case, right? What is the mean value?

Wolfgang.