The ‘evl’ command

The ‘evl’ umbrella utility can run the set of base commands available for controlling, inspecting and testing the state of the EVL core and any command matching the ‘evl-*’ glob pattern which may be reachable from the shell $PATH variable. The way the ‘evl’ utility centralizes access to a variety of EVL-related commands is very similar to that of git on purpose. Each of the EVL commands is implemented by an external plugin, which can be a mere executable, or a script in whatever language. The only requirement is that the caller must have execute permission on such file to run it.

The general syntax is as follows:

evl [-V] [-P <cmddir>] [-h] [<command> [command-args]]
  • <command> may be any command word listed by ‘evl -h’, such as: check which checks a kernel configuration for common issues ps which reports a snapshot of the current EVL threads test for running the EVL test suite trace which is a simple front-end to the ftrace interface for EVL

  • -P switches to a different installation path for base command plugins, which is located at $prefix/libexec by default.

  • -V displays the version information then exits. The information is extracted from the libevl library the EVL command depends on, displayed in the following format:

    evl.<serial> -- #<git-HEAD-commit> (<git-HEAD-date>) [ABI <revision>]

    where <serial> is the libevl serial release number, the <git-HEAD> information refers to the topmost GIT commit which is present in the binary distribution the ‘evl’ command is part of, and <revision> refers to the kernel ABI this binary distribution is compatible with. For instance:

    ~ # evl -V
    evl.0 -- #1c6115c (2020-03-06 16:24:00 +0100) [requires ABI 19]
    

The information following the double dash may be omitted if the built sources were not version-controlled by GIT.

  • given only -h or without any argument, ‘evl’ displays this general help, along with a short help string for each of the supported commands found in <cmddir>, such as:

    ~ # evl
    usage: evl [options] [<command> [<args>]]
    -P --prefix=<path>   set command path prefix
    -V --version         print library and required ABI versions
    -h --help            this help
    
    available commands:
    
    check        check kernel configuration
    gdb          debug EVL command plugin with GDB
    ps           report a snapshot of the current EVL threads
    test         run EVL tests
    trace        ftrace control front-end for EVL
    

Checking a kernel configuration (check)

evl check may be the very first evl command you should run from a newly installed target system which is going to run the EVL core. This command checks a kernel configuration for common issues which may increase latency. The general syntax is as follows:

$ evl check [-f --file=<.config>]
      	    [-L --check-list=<file>]
      	    [-a --arch=<cpuarch>]
	    [-H --hash-size=<N>]
	    [-q --quiet]
	    [-h --help]

The kernel configuration to verify is a regular .config file which contains all the settings for building a kernel image. If none is specified using the -f option, the command defaults to reading /proc/config.gz on the current machine. If this fails because any of CONFIG_IKCONFIG or CONFIG_IKCONFIG_PROC was disabled in the running kernel, the command fails.

The check list contains a series of single-line assertions which are tested against the contents of the kernel configuration. You can override the default check list stored at $prefix/libexec/kconf-checklist.evl with our own set of checks with the -L option. Each assertion follows the BNF-like syntax below:

assertion   : expr conditions
            | "!" expr conditions

expr        : symbol /* matches =y and =m */
            | symbol "=" tristate

tristate  : "y"
          | "m"
          | "n"

conditions  : dependency
            | dependency arch

dependency  : "if" symbol       /* true if set as y/m */

arch        : "on" cputype

cputype     : $(uname -m)

For instance:

  • CONFIG_FOO must be set whenever CONFIG_BAR is unset can be written as CONFIG_FOO if !CONFIG_BAR.

  • CONFIG_FOO must not be set can be written as !CONFIG_FOO, or conversely CONFIG_FOO=n.

  • CONFIG_FOO must be built as module on aarch32 or aarch64 can be written as CONFIG_FOO=m on aarch.

  • CONFIG_FOO must not be built-in on aarch64 if CONFIG_BAR is set can be written as !CONFIG_FOO=y if CONFIG_BAR on aarch.

Assertions in the check list may apply to a particular CPU architecture. Normally, the command should be able to figure out which architecture the kernel configuration file applies to by inspecting the first lines, looking for the “Linux/” pattern. However, you might have to specify this information manually to the command using the -a option if the file referred to by the -f option does not contain such information. The architecture name (cputype) should match the output of $(uname -m) or some abbreviated portion of it. However, arm64 and arm are automatically translated to aarch64 and aarch32 when found in an assertion or passed to the -a option.

The default check list translates the configuration-related information gathered in the caveat section as follows:

CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE=y if CONFIG_CPU_FREQ
CONFIG_DEBUG_HARD_LOCKS=n
CONFIG_ACPI_PROCESSOR_IDLE=n
CONFIG_LOCKDEP=n
CONFIG_DEBUG_LIST=n
CONFIG_DEBUG_VM=n
CONFIG_DEBUG_PER_CPU_MAPS=n
CONFIG_KASAN=n

The command returns the following information:

  • the wrong settings detected in the kernel configuration are written to stdout, unless the quiet -q option was given.

  • the number of failed assertions is returned via the shell exit code ($?).

Example: checking the current kernel configuration

~ # evl check
CONFIG_ACPI_PROCESSOR_IDLE=y
~ # echo $?
1

Reporting a snapshot of the current EVL threads (ps)

When you need to know which threads are currently present in your system, evl ps comes in handy. The command syntax - which supports short and long options formats - is as follows:

$ evl ps [-c --cpu=<cpu>[,<cpu>...]]
         [-s --state]
         [-t --times]
         [-p --policy]
         [-l --long]
         [-n --numeric]
         [-S --sort=<key>]
         [-h --help]

This command fetches the information it needs from the /sysfs attributes the EVL core exports for every thread it manages. The output is organized in groups of values, representing either runtime parameters or statistics for each displayed thread:

  • NAME reports the thread name, as specified in the evl_attach_self() library call.

  • CPU is the processor id. the thread is currently pinned to.

  • PID is the process id. inband-wise, since any EVL thread is originally a regular Linux [k]thread. This value belongs to the global namespace (i.e. task_pid_nr()).

  • SCHED is the current scheduling policy for the thread.

  • PRIO is the priority level in the scheduling policy for the thread.

  • ISW counts the number of inband switches. Under normal circumstances, this count should remain stable over time once the thread has entered its work loop. As the only exception, a thread which undergoes the SCHED_WEAK policy may see this counter progress as a result of calling out-of-band services. For all other scheduling policies, observing any increase in this value after the time-critical loop was entered is a sure sign of a problem in the application code, which might be calling real-time unsafe services when it should not.

  • CTXSW counts the number of context switches performed by the EVL core for the thread. This value is incremented each time the thread resumes from preemption or suspension in the EVL core. CAUTION: this has nothing to do with the context switches performed by the main kernel logic.

  • SYS is the number of out-of-band system calls issued by the thread to the EVL core. Basically, this is the number of out-of-band I/O requests the thread has issued so far, since everything is a file in the EVL core.

  • RWA counts the number of Remote WAkeup signals the EVL core had to send so far for waking up the thread whenever it was sleeping on a remote CPU. For instance, this would happen if two threads running on different CPUs were to synchronize on an EVL event. Waking up a remote thread entails sending an inter-processor interrupt to the CPU that thread sleeps on for kicking the rescheduling procedure, which entails more overhead than a local wakeup. If this counter increases like crazy when your application runs, you might want to check the situation with respect to CPU affinity, to make sure the current distribution of threads over the available CPUs is actually what you want.

  • STAT gives an abbreviated runtime status of the thread as follows:

    • ‘w’/‘W’ ⇾ Waits on a resource with/without timeout (TIMEOUT displays the time before timeout)
    • ’D’ ⇾ Delayed (TIMEOUT displays the remaining sleep time)
    • ‘p’ ⇾ Periodic timeline (kthread only, TIMEOUT displays the remaining time until the next period starts)
    • ‘R’ ⇾ Ready to run (i.e. neither blocked nor suspended, but waiting for the CPU to be available)
    • ‘X’ ⇾ Running in-band
    • ’T’ ⇾ Ptraced and stopped (whenever traced by a debugger such as gdb[server])
    • ‘r’ ⇾ Undergoes round-robin (when SCHED_RR is in effect)
    • ’S’ ⇾ Forcibly suspended (cumulative with ‘W’ state, won’t resume until lifted)
  • TIMEOUT is the remaining time before some timer which was started specifically for the thread fires. Which timer was started depends on the undergoing operation for such thread, which may block until a resource is available, wait for the next period in a timeline and so on. See STAT for details.

  • %CPU is the current amount of CPU horsepower consumed by the thread over the last second. When an out-of-band interrupt preempts a thread, the time spent handling it is charged to that thread.

  • CPUTIME reports the cumulated CPU time already consumed by the thread, using a minutes:milliseconds.microseconds format.

The command options allow to select which threads and which data should be displayed:

  • the -c option filters the output on the CPU the threads are pinned on. The argument is a comma-separated list of CPU numbers. Ranges are also supported via the usual dash separator. For instance, the following command would report threads pinned on CPU0, and all CPUs from CPU3 to CPU15.

    $ evl ps -c 0,3-15
    
  • -s includes information about the thread state, which is ISW, CTXSW, SYS, RWA and STAT.

  • -t includes the thread times, which are TIMEOUT, CPU% and CPUTIME.

  • -p includes the scheduling policy information, which is SCHED and PRIO.

  • -l enables the long output format, which is a combination of all information groups.

  • -n selects a numeric output for the STAT field, instead of the one-letter flags. This actually dumps the 32-bit value representing all aspects of a thread status in the EVL core, which contains more information than reported by the abbreviated format. EVL hackers want that.

  • -S sorts the output according to a given sort key in increasing order. The following sort keys are understood:

    • ‘c’ sorts by CPU values
    • ‘i’ sorts by ISW values
    • ’t’ sorts by CPUTIME values
    • ‘x’ sorts by CTXSW values
    • ‘w’ sorts by RWA values
    • ‘r’ sorts in reverse order (decreasing order)

For instance, the following command would list the times of all threads from CPU14 by decreasing CPU time consumption:

$ evl ps -t -Srt -c14
CPU   PID   TIMEOUT      %CPU   CPUTIME     NAME
 14   2603     -           0.6  00:435.952  rtup_ufpp14-5:2069
 14   2604     -           0.6  00:430.147  rtup_ufpp14-6:2069
 14   2599     -           0.5  00:423.118  rtup14-3:2069
 14   2600     -           0.5  00:420.293  rtup14-4:2069
 14   2595     -           0.3  00:207.143  [rtk1@14:2069]
 14   2597     -           0.3  00:204.301  [rtk2@14:2069]
 14   2619     -           0.2  00:186.139  rtuo_ufpp14-14:2069
 14   2617     -           0.2  00:185.497  rtuo_ufpp14-13:2069
 14   2623     -           0.2  00:184.812  rtuo_ufpp_ufps14-18:2069
 14   2622     -           0.2  00:184.772  rtuo_ufpp_ufps14-17:2069
 14   2621     -           0.2  00:181.692  rtuo_ufps14-16:2069
 14   2616     -           0.2  00:181.329  rtuo14-12:2069
 14   2615     -           0.2  00:181.230  rtuo14-11:2069
 14   2620     -           0.2  00:180.604  rtuo_ufps14-15:2069
 14   2572     -           0.0  00:000.006  rtus_ufps13-9:2069
 14   2125     -           0.0  00:000.006  rtus1-7:2069
 14   2646     -           0.0  00:000.005  rtus15-8:2069
 14   2650     -           0.0  00:000.005  rtus_ufps15-10:2069
 14   2310     -           0.0  00:000.005  rtus6-7:2069
  • -h displays the command help string.

Controlling the kernel tracer (trace)

The trace command provides a simple front-end for controlling the function tracer which is part of the FTRACE kernel framework, in a way which is Dovetail-aware. We typically use this tracer to analyze high latency spots during the course of the latmus program execution.

There is no Dovetail (or EVL-specific) tracer. Latency spots can be analyzed using the common kernel function tracer, which reports additional information about the current execution stage and interrupt state. Trace snapshots are automatically taken at appropriate times by the latmus utility in order to help in such analysis.

In order to use this tracer, make sure to enable the following features in your kernel build:

  • CONFIG_TRACER_SNAPSHOT
  • CONFIG_TRACER_SNAPSHOT_PER_CPU_SWAP
  • CONFIG_FUNCTION_TRACER

The command syntax is as follows:

evl trace [-e [-s <buffer_size>]]
    	  [-d] [-p] [-f] [-h]
	  [-c <cpu>]
	  [-h]

The command options allow for a straightforward use of the function tracer:

  • -e enables the tracer in the kernel. From this point, FTRACE starts logging information about a set of kernel functions which may be traversed while the system executes. By default, this switch only enables tracing for out-of-band IRQ events, CPU idling events, and all (in-kernel) EVL core routines. If a particular CPU is mentioned with -c along with -e, then per-CPU tracing is enabled for <cpu>.

  • if -f is mentioned, all kernel functions traversed in the course of execution are logged, not only the minimal subset enabled by default by -e. CAUTION: enabling full tracing may cause a massive overhead.

  • -s changes the size of the FTRACE buffer on each tracing CPU to <buffer_size>. If a particular CPU is mentioned with -c along with -s, then the change is applied to the snapshot buffer of <cpu> only.

  • -d fully disables the tracer which stops logging events on all CPUs.

  • -p prints out the contents of the trace buffer. If a particular CPU is mentioned with -c along with -p, then only the snapshot buffer of <cpu> is dumped.

  • -h displays the command help string.

For instance, the following command starts tracing all kernel routines:

$ evl trace -ef

Interpreting the Dovetail-specific trace information

The Dovetail-specific information is about:

  • whether the in-band stage is stalled and/or irqs are disabled in the CPU. ’d’ appears in the entry state flags if the in-band stage is stalled while hard irqs are enabled in the CPU, ’D’ denotes an unstalled in-band stage with hard irqs off in the CPU, and ‘*’ denotes a combined stalled in-band stage and hard irqs off in the CPU.

  • whether we are running on the out-of-band stage, if ‘~’ appears in the entry flags.

You may want to read this document for details on the notion of interrupt stage Dovetail implements.

For instance:

/* hard irqs off, running in-band */
<...>-4164  [003] D...   122.047972: do_syscall_64 <-entry_SYSCALL_64_after_hwframe
/* in-band stalled and hard irqs off, running out-of-band */
<...>-4164  [003] *.~.   122.048021: __evl_schedule <-run_oob_call
/* in-band stalled, hard irqs on, running in-band */
<...>-4164  [003] d...   122.048082: rcu_lockdep_current_cpu_online <-rcu_read_lock_sched_held

In addition to this basic information, latmus emits a special tracepoint named evl_latspot in the trace event log before taking a trace snapshot, each time the observed maximum latency increases. The frozen trace is visible in the corresponding per-CPU snapshot buffer. From that point, you may be able to backtrack to the source(s) of the extra latency. A typical debug session would look like this:

~ # evl trace -ef
tracing enabled

~ # latmus
warming up on CPU1...
RTT|  00:00:01  (user, 1000 us period, priority 98, CPU1)
RTH|----lat min|----lat avg|----lat max|-overrun|---msw|---lat best|--lat worst
RTD|     26.675|     26.951|     27.826|       0|     0|     26.675|     27.826
RTD|     26.712|     27.067|     31.204|       0|     0|     26.675|     31.204
RTD|     26.653|     26.961|     29.160|       0|     0|     26.653|     31.204
RTD|     26.678|     27.067|     29.285|       0|     0|     26.653|     31.204
RTD|     26.759|     27.051|     29.542|       0|     0|     26.653|     31.204
RTD|     26.770|     27.079|     29.266|       0|     0|     26.653|     31.204
^C---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     10.119|     27.029|     31.204|       0|     0|    00:00:06/00:00:06

~ # evl trace -c 1
 ...
          <idle>-0     [001] *.~.   135.363256: do_trace_write_msr <-__switch_to
          <idle>-0     [001] *.~.   135.363256: write_msr: c0000100, value 7ff90973e700
 timer-responder-234   [001] *.~.   135.363256: switch_fpu_return <-dovetail_context_switch
 timer-responder-234   [001] *.~.   135.363257: do_raw_spin_unlock <-__evl_schedule
 timer-responder-234   [001] *.~.   135.363257: do_raw_spin_lock <-evl_wait_schedule
 timer-responder-234   [001] *.~.   135.363258: do_raw_spin_unlock <-evl_wait_schedule
 timer-responder-234   [001] *.~.   135.363258: do_raw_spin_lock <-latmus_oob_ioctl
 timer-responder-234   [001] *.~.   135.363258: do_raw_spin_unlock <-latmus_oob_ioctl
 timer-responder-234   [001] d.~.   135.363259: evl_oob_sysexit: result=0
 timer-responder-234   [001] d.~.   135.363262: pipeline_syscall <-do_syscall_64
 timer-responder-234   [001] d.~.   135.363262: handle_oob_syscall <-pipeline_syscall
 timer-responder-234   [001] d.~.   135.363263: do_oob_syscall <-handle_oob_syscall
 timer-responder-234   [001] d.~.   135.363263: evl_oob_sysentry: syscall=oob_ioctl
 timer-responder-234   [001] d.~.   135.363264: EVL_ioctl <-do_oob_syscall
 timer-responder-234   [001] d.~.   135.363264: evl_get_file <-EVL_ioctl
 timer-responder-234   [001] *.~.   135.363264: do_raw_spin_lock <-evl_get_file
 timer-responder-234   [001] *.~.   135.363265: do_raw_spin_unlock <-evl_get_file
 timer-responder-234   [001] d.~.   135.363265: latmus_oob_ioctl <-EVL_ioctl
 timer-responder-234   [001] d.~.   135.363266: add_measurement_sample <-latmus_oob_ioctl
 timer-responder-234   [001] d.~.   135.363266: evl_latspot: ** latency peak: 31.204 us **

In the latmus case, part of this analysis would include estimating the delay between the latest tick date programmed in the hardware and the actual receipt of the timer interrupt. When tracing is enabled, this information is automatically produced in the trace log:

/* This is when the timer chip is programmed for the next tick. */
<idle>-0     [001] *.~.   135.362244: evl_timer_shot: latmus_pulse_handler at 135.363228 (delay: 984 us, 196195 cycles
...
/* This is when the corresponding timer interrupt is received by Dovetail. */
<idle>-0     [001] *.~.   135.363233: irq_handler_entry: irq=4354 name=Out-of-band LAPIC timer interrupt

Running the test suite (test)

This command is a short-hand for running the EVL test suite. The usage is as follows:

$ evl test [-l][-L][-k] [test-list]

With no argument, this command runs all of the tests available from the default installation path at $prefix/tests:

$ evl test
duplicate-element: OK
monitor-pp-dynamic: OK
monitor-pi: OK
clone-fork-exec: OK
...

You can also chose to run a specific set of tests by mentioning them as arguments to the command, such as:

$ evl test duplicate-element monitor-pi
duplicate-element: OK
monitor-pi: OK

You may ask for listing the available tests instead of executing them, by using the -l switch:

$ evl test -l
duplicate-element
monitor-pp-dynamic
monitor-pi
clone-fork-exec
...

In a variant aimed at making scripting easier, you can ask for the absolute paths instead:

$ evl test -L proxy-pipe mapfd
/usr/evl/tests/proxy-pipe
/usr/evl/tests/mapfd

If some test goes wrong, the command normally stops immediately. Passing -k would allow it to keep going until the end of the series.

Implementing your own EVL commands

You can implement your own ‘evl’ command plugins, which may be located anywhere provided it is reachable from the shell PATH variable with the proper execute permission bit set. EVL comes with a set of base plugins available from $prefix/libexec (*). The latter directory is implicitly searched for the command after the PATH variable was considered, which means that you may override any base command with your own implementation whenever you see fit.

Crash course: adding the ‘foo’ command script to ~/tools

$ mkdir ~/tools
$ cat > ~/tools/evl-foo
#! /bin/sh
echo "this is your 'evl foo' command"
^D
$ chmod +x ~/tools/evl-foo
$ export PATH=$PATH:~/tools
$ evl foo
this is your 'evl foo' command

In addition, ‘evl’ sets a few environment variables before calling a plugin. Your plugin executable/script can retrieve them using getenv(3) from a C program, or directly dereference those variables from a shell:

Variable Description Default value
EVL_CMDDIR Where to find the base plugins $prefix/libexec
EVL_TESTDIR Where to find the tests $prefix/tests
EVL_SYSDIR root of the /sysfs hierarchy for EVL devices /sys/devices/virtual
EVL_TRACEDIR root of ftrace hierarchy /sys/kernel/debug/tracing

(*) may be overriden using the -P option.


Last modified: Wed, 12 Aug 2020 18:12:06 CEST