Re: hith latency when cpu is fully loaded

[chensong <chensong@kylinos.cn>]

On 2019年11月19日 14:49, Jan Kiszka wrote:
> On 19.11.19 02:01, chensong via Xenomai wrote:
>> Dear experts,
>>
>> i'm new in xenomai, i got an issue, here is the detail:
>>
>> Main processor architect: ARM64 phytium ft2000ahk
>> Kernel release number: 4.14.4
>> cmdline:BOOT_IMAGE=/Image-tmp
>> root=UUID=9fea0634-a9c9-4e9f-906c-9c36b7249822 console=ttyS1,115200
>> earlyprintk=uart8250-32bit,0x28001000 rw rootdelay=10 KEYBOARDTYPE=pc
>> KEYTABLE=us security=
>> xenomai release number:3.1-devel
>> xenomai configuration:
>>         kylin@kylin-os:~/workspace/code/nudt-hgj-xenomai-tjrd$ grep
>> configure config.status
>>         # Generated by configure.
>>        # Compiler output produced by configure, useful for debugging
>>         # configure, is in config.log if it exists.
>>         configured by ./configure, generated by GNU Autoconf 2.69,
>>        ac_configure_extra_args=
>>        ac_configure_extra_args="$ac_configure_extra_args --silent"
>>        set X /bin/bash './configure'  '--with-core=cobalt'
>> '--enable-smp' '--enable-pshared' $ac_configure_extra_args --no-create
>> --no-recursion
>>        configure_time_dlsearch_path='/lib /usr/lib
>> /lib/aarch64-linux-gnu /usr/lib/aarch64-linux-gnu
>> /usr/lib/aarch64-linux-gnu/mesa-egl /usr/lib/aarch64-linux-gnu/mesa
>> /usr/local/lib '
>>        configure_time_lt_sys_library_path=''
>>       for var in reload_cmds old_postinstall_cmds
>> old_postuninstall_cmds old_archive_cmds extract_expsyms_cmds
>> old_archive_from_new_cmds old_archive_from_expsyms_cmds archive_cmds
>> archive_expsym_cmds module_cmds module_expsym_cmds export_symbols_cmds
>> prelink_cmds postlink_cmds postinstall_cmds postuninstall_cmds
>> finish_cmds sys_lib_search_path_spec configure_time_dlsearch_path
>> configure_time_lt_sys_library_path; do
>>        # on some systems where configure will not decide to define it.
>>        # Let's still pretend it is `configure' which instantiates
>> (i.e., don't
>>        configure_input='Generated from '`
>>         `' by configure.'
>>        configure_input="$ac_file.  $configure_input"
>>         case $configure_input in #(
>>         ac_sed_conf_input=`$as_echo "$configure_input" |
>>         *) ac_sed_conf_input=$configure_input;;
>>          s|@configure_input@|$ac_sed_conf_input|;t t
>>        $as_echo "/* $configure_input  */" \
>>      $as_echo "/* $configure_input  */" \
>>       # Libtool was configured on host `(hostname || uname -n)
>> 2>/dev/null | sed 1q`:
>>        : \${LT_SYS_LIBRARY_PATH="$configure_time_lt_sys_library_path"}
>>         sys_lib_dlsearch_path_spec=$lt_configure_time_dlsearch_path
>>        # Explicit LT_SYS_LIBRARY_PATH set during ./configure time.
>> configure_time_lt_sys_library_path=$lt_configure_time_lt_sys_library_path
>>
>> OR:
>>         kylin@kylin-os:~$ xeno-config --info
>>          Xenomai version: Xenomai/cobalt v3.1-devel -- # ()
>>         Linux kylin-os 4.14.4.kylin.rt-1118-ipipe-trace+ #2 SMP
>> PREEMPT Mon Nov 18 18:28:17 CST 2019 aarch64 aarch64 aarch64 GNU/Linux
>>         Kernel parameters: BOOT_IMAGE=/Image-tmp
>> root=UUID=9fea0634-a9c9-4e9f-906c-9c36b7249822 console=ttyS1,115200
>> earlyprintk=uart8250-32bit,0x28001000 rw rootdelay=10 KEYBOARDTYPE=pc
>> KEYTABLE=us security=
>>         I-pipe release #2 detected
>>         Cobalt core 3.1-devel detected
>>         Compiler: gcc version 5.4.0 20160609 (Ubuntu/Linaro
>> 5.4.0-6kord1~16.04.10)
>>         Build args: --prefix=/usr --includedir=/usr/include/xenomai
>> --mandir=/usr/share/man --with-testdir=/usr/lib/xenomai/testsuite
>> --enable-smp --build aarch64-linux-gnu build_alias=aarch64-linux-gnu
>>
>> Desktop: kylin 4.0.2 (ubuntu likely desktop)
>>
>> Issue description:
>>        latency and cyclictest work fine in my system in most of cases,
>> the worst latency is around 100us ~ 200us. however, when i ran a
>> script to increase the cpu load in the system, the worst latency
>> reached 2000us ~ 5000us or even worse. Basically, the script forks 6
>> processes by default and each process applies a four-pages buffer and
>> keeps writing without any breath, no warning or error messages in
>> dmesg. below is the script:
>>
>> #include <stdio.h>
>> #include <stdlib.h>
>> #include <sys/types.h>
>> #include <sys/wait.h>
>> #include <sys/stat.h>
>> #include <fcntl.h>
>>
>> #define PAGE_SIZE 4096
>> #define TEST_THREADS 6
>>
>> unsigned int test_threads;
>>
>> void do_thread_test(void)
>> {
>>      void *mm;
>>      char i = 0;
>>
>>      printf("mem test thread start \n");
>>      mm = malloc(PAGE_SIZE * 4); // 1M
>>      while(1) {
>>          for (i = 0; i<100; i++)
>>              memset(mm, i, PAGE_SIZE * 4);
>>      }
>
> You cannot run Xenomai threads at 100% on Linux. You need to leave some
> time for the rest of the system to do housekeeping. That explains the
> "deadlock" you see. If you turn on CONFIG_XENO_OPT_WATCHDOG, it will
> detect such mistakes and kick the task out of RT.
>
> Jan
>

"do_thread_test" is not an RT task, it's running in linux domain and 
latency running in xenomai domain was affected.

/Song