Add bpftrace to pbench-agent-tools

[sourabhtk37]

Todo

• find where bpftrace is installed: /usr/bin/ or /usr/local/bin
• Check for required packages and kernel headers for bpftrace to run

---

This is for tracking purpose. I used the existing jmap tool to base this. Need your inputs/reviews on implementing the post-processing and other features needed for pbench.

Here's the sample stdout from a run, which uses a basic bpftrace program, syscount.bt:

# cat /var/lib/pbench-agent/iozone__2019.02.26T16.23.27/1/reference-result/tools-default/localhost/bpftrace/bpftrace-stdout.txt
Attaching 3 probes...
Counting syscalls... Hit Ctrl-C to end.
 
Top 10 syscalls IDs:
@syscall[202]: 82215
@syscall[186]: 157832
@syscall[232]: 176516
@syscall[38]: 183350
@syscall[7]: 228544
@syscall[20]: 253303
@syscall[8]: 284114
@syscall[1]: 311635
@syscall[0]: 465967
@syscall[47]: 682795
 
Top 10 processes:
@process[grep]: 72991
@process[cut]: 75418
@process[Chrome_IOThread]: 89515
@process[Chrome_ChildIOT]: 100151
@process[code]: 167427
@process[sh]: 358588
@process[chromium-browse]: 365207
@process[xprop]: 622420
@process[Xorg]: 669920
@process[iozone]: 921651

Signed-off-by: T K Sourab sourabhtk37@gmail.com

[atheurer]

1. Can you provide examples of how the bpftrace tool was registered and how you started/stopped the tool (using a pbench benchmark or running pbench-start-tools or pbench-stop-tools)?

2. Does syscount.bt require compiling before using? If so, please see how this is handled for systemtap for pbench (the stap code is compiled during --install, which what pbench-register-tool calls), as bpf might need to do this in a similar way.

3. The --install will need to check for required packages (not cmds) and install if not found (using check_install_rpm)

[portante]

Great start. This is really cool.

Can you be sure to follow the model of the systemtap tool as well?

[sourabhtk37]

@atheurer 2. The syscount.bt doesn't need compiling into the kernel, bpftrace using libbpf dynamically loads it into the kernel.
3. Yes, I'm working on that. Need to make sure kernel configs required to run is enabled and other packages required for running are also installed. Thanks.

[atheurer]

@atheurer 2. The syscount.bt doesn't need compiling into the kernel, bpftrace using libbpf dynamically loads it into the kernel.

Yes, I understand it can do this, but my concern is that libbpf doing this at tool-start time might require a lot of [cpu] resources, or might require a little bit of time to do this. As a comparison, for systemtap, this is why we "pre" compile the script when we register the tool, instead of just doing it automatically at start-tools. Maybe this is not possible with BPF, but I guess we can see how it behaves with the way you have it now, and then see later if it becomes an issue.

[portante]

Almost, just a few things to clean up.

[portante]

Almost there, looking good.

[portante]

Can you also rebase and squash this commits?

[portante]

Looks good, tried it out with a mock of /usr/bin/bpftrace and it seemed to work fine.