Exercise 6

This exercise is not so different than any latency measurement script where latency is measured as difference between timestamps of two probe firings and saved to an aggregation.

Note that plockstat$ provider doesn't serve a probe for mutex lock attempt, so we had to expand it by using pid$ provider. As you may notice from ustack outputs in pthread.d example, mutex lock attempts are implemented by mutex_lock_impl() libc function. We will use quantize() aggregation which will be printed with printa:

  Script file mtxtime.d

    self->mtxtime = timestamp; 

/ self->mtxtime != 0 /
    @[ustack()] = quantize(timestamp - self->mtxtime);
    self->mtxtime = 0;


You will need to bind this script to tsexperiment process using -c or -p option.

We will need to use static probes mutex_entry and mutex_acquired for a SystemTap version of that script. However, we will need to be careful while working with userspace backtraces. First of all, we should use -d option to provide path to SystemTap for resolving symbols or --ldd to make it scan library dependencies of traced binary and automatically add them (when some of them are missing, stap utility will provide a hint with full paths).

Mutexes are also often used in TSLoad which can cause excessive overheads when we try to trace them, especially when we will use ubacktrace() function. You can use STP_NO_OVERLOAD macro definition (which can be passed to stap with -D option) to prevent stap from failing when overheads are big, or you can reduce overheads. In our case we will limit amount of traced callers by using ucallers() function which accepts depth of backtrace as a first argument like ustack() function from backtrace and only collects addresses without resolving them to symbols. We will defer symbol resolving to an aggregation printing.

Here are our script for SystemTap:

  Script file mtxtime.stp

global mtxtime[128], mtxlockt;

@define libpthread %( "/lib64/libpthread.so.0" %)
@define tsexperiment %( "/opt/tsload/bin/tsexperiment" %)

probe process(@libpthread).mark("mutex_entry") {
    if(pid() != target()) next;

    mtxtime[tid()] = local_clock_ns();

probe process(@libpthread).mark("mutex_acquired") {
    if(pid() != target()) next;

    tm = mtxtime[tid()];
    if(tm == 0) next;

    mtxlockt[ucallers(6)] <<< local_clock_ns() - tm;
    delete mtxtime[tid()];

probe process(@tsexperiment).function("experiment_unconfigure") {
    foreach([ub] in mtxlockt) {
        if(@count(mtxlockt[ub]) < 100)