Re: [lwip-users] Multi-Threaded netconn API TCP application

[Darius Babrauskas]

> Hi,
>
> On Thu, Oct 01, 2015 at 04:41:34PM +0300, Darius Babrauskas wrote:
> > Maybe, not was supported the Linux port. Strange, that gettimeofday
> > CLOCK_REALTIME was used long time.
>
> I don't understand any of that.
>
>
> > This can damage the test results. Exmp. NTP client can change
> > systemtime at each half hour, some seconds, when your testing LWIP.

    linux have adjtime, which using ntp client for "slew " when time 
different is small < 128ms.
    If bigger different, doing "steps" (maybe small < 1000s)
    This from my Ubuntu whireshark log (I think that wireshark using 
gettimeofday CLOCK_REALTIME ):
No.     Time        Source                Destination           Protocol 
Length Info
     1 0.000000    10.0.1.158            91.189.94.4           NTP      90 
NTP Version 4, client
     2 0.030025    10.0.1.158            10.0.1.238            X11      686 
Requests: CreatePixmap, <Unknown opcode 145>,
     3 0.041885    ViaTechn_              Broadcast             ARP      60 
Who has 10.0.1.158?  Tell 10.0.1.254
     4 0.041896    AsustekC_             ViaTechn_f1           ARP      42 
10.0.1.158 is at f4:6d:.......
     5 0.042006    91.189.94.4           10.0.1.158            NTP      90 
NTP Version 4, server
     6 -0.279864   10.0.1.158            10.0.1.238            X11      686 
Requests: CreatePixmap, <Unknown opcode 145>, <Unknown opcode 145>,
     7 -0.279646   10.0.1.238            10.0.1.158            TCP      66 
6000→47917 [ACK] Seq=1 Ack=1241 Win=65535 Len=0 TSval=301879 TSecr=7480572

> No, that's not how NTP works on Linux. NTP does not step the system
> time, it makes very small adjustments in the system clock frequency in
> order to slow down or hurry up very very smoothly the system time.
> CLOCK_MONOTONIC is still affected by incremental NTP adjustments.
But it  "slew ". I think using CLOCK_MONOTONIC are better than 
CLOCK_MONOTONIC_RAW.


> If you are talking about a NTP client called from time to time to set
> the system clock, e.g: by calling ntpdate in a cronjob, then you are
> doing it wrong, terribly wrong.
Yes, here more problems, which need resolve.

> Embedded devices are of different concern, NTP is generally not used to
> change a clock input used for timers or scheduler and very few devices
> are adjusting a timer frequency based on the NTP timing data. Therefore
> it is almost only used to output the current date in logs or human
> interfaces, thus we don't care if it goes backward or forward a little
> bit.
In  my "embedded" projects I using SNTP for time synchronization without big 
problems, because use CLOCK_MONOTONIC.

> Sylvain

Darius