On Mar 2, 2020, at 5:36 AM, Quentin Monnet [off-list ref] wrote:

Hi Song,

Thanks for this work! Some remarks (mostly nitpicks) inline.

2020-02-28 15:40 UTC-0800 ~ Song Liu [off-list ref]

quoted

With fentry/fexit programs, it is possible to profile BPF program with
hardware counters. Introduce bpftool "prog profile", which measures key
metrics of a BPF program.

bpftool prog profile command creates per-cpu perf events. Then it attaches
fentry/fexit programs to the target BPF program. The fentry program saves
perf event value to a map. The fexit program reads the perf event again,
and calculates the difference, which is the instructions/cycles used by
the target program.

Example input and output:

 ./bpftool prog profile 3 id 337 cycles instructions llc_misses

       4228 run_cnt
    3403698 cycles                                              (84.08%)
    3525294 instructions   #  1.04 insn per cycle               (84.05%)
         13 llc_misses     #  3.69 LLC misses per million isns  (83.50%)

This command measures cycles and instructions for BPF program with id
337 for 3 seconds. The program has triggered 4228 times. The rest of the
output is similar to perf-stat. In this example, the counters were only
counting ~84% of the time because of time multiplexing of perf counters.

Note that, this approach measures cycles and instructions in very small
increments. So the fentry/fexit programs introduce noticeable errors to
the measurement results.

The fentry/fexit programs are generated with BPF skeletons. Therefore, we
build bpftool twice. The first time _bpftool is built without skeletons.
Then, _bpftool is used to generate the skeletons. The second time, bpftool
is built with skeletons.

Signed-off-by: Song Liu <redacted>
---

Thanks Yonghong and Quentin! I will fix these and send v3. 

Song