• Viktor Rosendahl's avatar
    tracing/tools: Add the latency-collector to tools directory · e23db805
    Viktor Rosendahl authored
    This is a tool that is intended to work around the fact that the
    preemptoff, irqsoff, and preemptirqsoff tracers only work in
    overwrite mode. The idea is to act randomly in such a way that we
    do not systematically lose any latencies, so that if enough testing
    is done, all latencies will be captured. If the same burst of
    latencies is repeated, then sooner or later we will have captured all
    the latencies.
    
    It also works with the wakeup_dl, wakeup_rt, and wakeup tracers.
    However, in that case it is probably not useful to use the random
    sleep functionality.
    
    The reason why it may be desirable to catch all latencies with a long
    test campaign is that for some organizations, it's necessary to test
    the kernel in the field and not practical for developers to work
    iteratively with field testers. Because of cost and project schedules
    it is not possible to start a new test campaign every time a latency
    problem has been fixed.
    
    It uses inotify to detect changes to /sys/kernel/tracing/trace.
    When a latency is detected, it will either sleep or print
    immediately, depending on a function that act as an unfair coin
    toss.
    
    If immediate print is chosen, it means that we open
    /sys/kernel/tracing/trace and thereby cause a blackout period
    that will hide any subsequent latencies.
    
    If sleep is chosen, it means that we wait before opening
    /sys/kernel/tracing/trace, by default for 1000 ms, to see if
    there is another latency during this period. If there is, then we will
    lose the previous latency. The coin will be tossed again with a
    different probability, and we will either print the new latency, or
    possibly a subsequent one.
    
    The probability for the unfair coin toss is chosen so that there
    is equal probability to obtain any of the latencies in a burst.
    However, this assumes that we make an assumption of how many
    latencies there can be. By default  the program assumes that there
    are no more than 2 latencies in a burst, the probability of immediate
    printout will be:
    
    1/2 and 1
    
    Thus, the probability of getting each of the two latencies will be 1/2.
    
    If we ever find that there is more than one latency in a series,
    meaning that we reach the probability of 1, then the table will be
    expanded to:
    
    1/3, 1/2, and 1
    
    Thus, we assume that there are no more than three latencies and each
    with a probability of 1/3 of being captured. If the probability of 1
    is reached in the new table, that is we see more than two closely
    occurring latencies, then the table will again be extended, and so
    on.
    
    On my systems, it seems like this scheme works fairly well, as
    long as the latencies we trace are long enough, 300 us seems to be
    enough. This userspace program receive the inotify event at the end
    of a latency, and it has time until the end of the next latency
    to react, that is to open /sys/kernel/tracing/trace. Thus,
    if we trace latencies that are >300 us, then we have at least 300 us
    to react.
    
    The minimum latency will of course not be 300 us on all systems, it
    will depend on the hardware, kernel version, workload and
    configuration.
    
    Example usage:
    
    In one shell, give the following command:
    sudo latency-collector -rvv -t preemptirqsoff -s 2000 -a 3
    
    This will trace latencies > 2000us with the preemptirqsoff tracer,
    using random sleep with maximum verbosity, with a probability
    table initialized to a size of 3.
    
    In another shell, generate a few bursts of latencies:
    
    root@host:~# modprobe preemptirq_delay_test delay=3000 test_mode=alternate
    burst_size=3
    root@host:~# echo 1 > /sys/kernel/preemptirq_delay_test/trigger
    root@host:~# echo 1 > /sys/kernel/preemptirq_delay_test/trigger
    root@host:~# echo 1 > /sys/kernel/preemptirq_delay_test/trigger
    root@host:~# echo 1 > /sys/kernel/preemptirq_delay_test/trigger
    
    If all goes well, you should be getting stack traces that shows
    all the different latencies, i.e. you should see all the three
    functions preemptirqtest_0, preemptirqtest_1, preemptirqtest_2 in the
    stack traces.
    
    Link: https://lkml.kernel.org/r/20210212134421.172750-2-Viktor.Rosendahl@bmw.deSigned-off-by: default avatarViktor Rosendahl <Viktor.Rosendahl@bmw.de>
    Signed-off-by: default avatarSteven Rostedt (VMware) <rostedt@goodmis.org>
    e23db805
Makefile 6.2 KB