[LTP] [PATCH v4 0/4] Basic eBPF tests

[Clemens Famulla-Conrad]

On Mon, 2019-08-26 at 10:29 -0400, Jan Stancek wrote:
> 
> ----- Original Message -----
> > I've ended up playing with the patchset and fixed a few loose ends
> > on
> > the map test and as I had the code at hand I decided to send v4
> > instead
> > of pointing out the mistakes in a review.
> > 
> > There were numerous small changes for the map test:
> > 
> > * Make sure the key buffer is sized exactly for the content
> > 
> > * Initialized the array/hash element value in test setup
> > 
> > * Made the code flow a bit more obvious, it was hard to tell which
> >   part was run for n == 0 and which for n == 1
> > 
> > Also it looks that for me the test that loads the eBPF program ends
> > up
> > with EPERM randomly at about 10th iteration both as unpriviledged
> > and
> > priviledged user, which is really strange.
> 
> There's one EPERM I can reproduce reliably with bpf_map test, which
> appears
> to originate from "bpf_charge_memlock".
> 
> There's a deferred component to map freeing, and unchange appears to
> be part of it:
>   bpf_map_release
>     bpf_map_put
>       INIT_WORK(&map->work, bpf_map_free_deferred);
>         (deferred) bpf_uncharge_memlock
> 
> When I lower max locked memory, it's easy to hit:
> # ulimit  -l 128; ./bpf_map01 -i 100
> ...
> bpf_map01.c:52: CONF: bpf() requires CAP_SYS_ADMIN on this system:
> EPERM
> 
> Can you try bumping max locked memory to some high value and check
> if that helps your case?

# for i in 64 128 256 1024; do
    echo $i; 
    ulimit -l $i;
    ./bpf_prog01 -i 100 |& grep -P 'passed|CONF'; 
done

64
CONF: bpf() requires CAP_SYS_ADMIN on this system: EPERM
passed   16

128
CONF: bpf() requires CAP_SYS_ADMIN on this system: EPERM
passed   16

256
CONF: bpf() requires CAP_SYS_ADMIN on this system: EPERM
passed   32

1024
CONF: bpf() requires CAP_SYS_ADMIN on this system: EPERM
passed   192


Which produce almost the same results.
Same approach with `bpf_map01` differs a lot. Sometimes all pass,
sometimes none.


/Clemens