Re: Using stapdyn to probe child processes of the target?

[Arkady <arkady dot miasnikov at gmail dot com>]

I have never tried user space hooking. My immediate idea is to run gcc
under strace, filter all fork-exec pairs (a Python script here?) and
attach stap to all child PIDs

On Fri, Jun 16, 2017 at 12:47 AM, Serhei Makarov <serhei.etc@gmail.com> wrote:
> Hello all,
>
> I've been trying to observe gcc's execution using Stapdyn. Most of the
> actual work is done by a child process (cc1) forked from gcc. (Below
> I've attached a script, proctree.stp, for listing processes forked by
> a target command.) If I'm using kernel systemtap, I can just target
> cc1 directly, for example:
>
> stap -ve 'probe
> process("/usr/libexec/gcc/x86_64-redhat-linux/6.3.1/cc1").insn {
> counter<<<1 } global counter probe end { printf("%d calls\n",
> @count(counter)) }' -c "gcc test/widget3.c"
>
> Instrumenting using kernel uprobes introduces quite a large
> performance overhead, so I would prefer Dyninst for this. But the same
> script doesn't work with --dyninst option, because the target process
> selected using -c is not cc1.
>
> As far as I can tell, my only options for observing cc1 with Stapdyn are:
>
> (a) use -c option : invoke cc1 directly. Feasible in this case, may
> not be feasible in other software where the parent/child processes
> coordinate using more complex IPC. This also doesn't give a profile of
> the full GCC execution (unlike, say, "valgrind --trace-children=yes").
> (b) use -x option : attach to cc1 by PID after it starts running. Not
> feasible in this case since cc1 is a batch program, not a long-running
> daemon.
>
> Am I missing any possibilities? Is it currently possible to use
> Stapdyn to attach to processes other than a single target?
>
> All the best,
>      Serhei Makarov
>
> PS Here's the proctree.stp script:
>
> global pids
>
> probe begin {
>   pids[target()] = 0
> }
>
> probe process.begin {
>   child = pid(); parent = ppid()
>   if (parent in pids) {
>     printf("fork %s(%d) -> %s(%d)\n", pexecname(), parent, execname(), child)
>     pids[child] = parent
>   }
> }
>
> This outputs things like:
>
> [serhei@beatrice e-tracer]$ stap proctree.stp -c "gcc test/widget3.c"
> Missing separate debuginfos, use: debuginfo-install
> kernel-core-4.10.10-200.fc25.x86_64
> fork gcc(2725) -> gcc(2728)
> fork gcc(2725) -> cc1(2728)
> fork gcc(2725) -> gcc(2729)
> fork gcc(2725) -> as(2729)
> fork gcc(2725) -> gcc(2730)
> fork gcc(2725) -> collect2(2730)
> fork collect2(2730) -> collect2(2731)
> fork collect2(2730) -> ld(2731)
>
> Or, here's a Firefox browsing session:
>
> [serhei@beatrice e-tracer]$ stap proctree.stp -c "firefox --new-instance"
> fork firefox(3179) -> firefox(3182)
> fork firefox(3179) -> basename(3182)
> fork firefox(3179) -> firefox(3183)
> fork firefox(3179) -> uname(3183)
> fork firefox(3179) -> firefox(3184)
> fork firefox(3179) -> pidof(3184)
> fork firefox(3179) -> firefox(3185)
> fork firefox(3179) -> mkdir(3185)
> fork firefox(3179) -> firefox(3186)
> fork firefox(3186) -> firefox(3187)
> fork firefox(3186) -> firefox(3188)
> fork firefox(3186) -> sed(3188)
> fork firefox(3179) -> firefox(3189)
> fork firefox(3189) -> firefox(3190)
> fork firefox(3189) -> firefox(3191)
> fork firefox(3189) -> sed(3191)
> fork firefox(3179) -> firefox(3192)
> fork firefox(3179) -> expr(3192)
> fork run-mozilla.sh(3179) -> run-mozilla.sh(3193)
> fork run-mozilla.sh(3179) -> basename(3193)
> fork run-mozilla.sh(3179) -> run-mozilla.sh(3194)
> fork run-mozilla.sh(3179) -> dirname(3194)
> fork run-mozilla.sh(3179) -> run-mozilla.sh(3195)
> fork run-mozilla.sh(3179) -> uname(3195)
> fork firefox(3179) -> firefox(3196)
> fork firefox(3179) -> firefox(3198)
> fork Gecko_IOThread(3179) -> Gecko_IOThread(3266)
> fork Gecko_IOThread(3179) -> firefox(3266)