Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20200522003850.GA32698@paulmck-ThinkPad-P72> <20200522094407.GK325280@hirez.programming.kicks-ass.net>
 <20200522143201.GB32434@rowland.harvard.edu> <20200522174352.GJ2869@paulmck-ThinkPad-P72>
 <006e2bc6-7516-1584-3d8c-e253211c157e@fb.com> <ac799c98-45dd-d056-386f-cbebc7270c0c@gmail.com>
 <CAEf4BzYGbLfGA=NN=dP1RqDj7yp_Fnu0L-1bgQojPMt0-Y_X=g@mail.gmail.com>
 <69ed3604-4275-d73e-a5d6-2b70dd877104@gmail.com> <CAEf4BzajE6jCkbBQ+f0cG=Y+vAEPWGNhfOMFVVhoDZWjNV-oGA@mail.gmail.com>
 <d1113b47-a920-c0e4-9aa4-88781368a26f@gmail.com> <b92f9c0a-826c-d074-5389-8c340f7cccf2@gmail.com>
In-Reply-To: <b92f9c0a-826c-d074-5389-8c340f7cccf2@gmail.com>
From:   Andrii Nakryiko <andrii.nakryiko@gmail.com>
Date:   Tue, 26 May 2020 13:19:36 -0700
Message-ID: <CAEf4BzZF6JO9Tpc2wRk1GDHKfDr0LeYz7LrHz93W1uNjSaq=bg@mail.gmail.com>
Subject: Re: Some -serious- BPF-related litmus tests
To:     Akira Yokosawa <akiyks@gmail.com>
Cc:     luc.maranget@inria.fr, Andrii Nakryiko <andriin@fb.com>,
        "Paul E . McKenney" <paulmck@kernel.org>,
        Alan Stern <stern@rowland.harvard.edu>,
        Peter Zijlstra <peterz@infradead.org>, parri.andrea@gmail.com,
        will@kernel.org, Boqun Feng <boqun.feng@gmail.com>,
        npiggin@gmail.com, dhowells@redhat.com, j.alglave@ucl.ac.uk,
        dlustig@nvidia.com, Joel Fernandes <joel@joelfernandes.org>,
        open list <linux-kernel@vger.kernel.org>,
        linux-arch@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, May 26, 2020 at 7:02 AM Akira Yokosawa <akiyks@gmail.com> wrote:
>
> On Tue, 26 May 2020 19:50:47 +0900, Akira Yokosawa wrote:
> > On Mon, 25 May 2020 16:31:05 -0700, Andrii Nakryiko wrote:
> >> On Mon, May 25, 2020 at 3:01 PM Akira Yokosawa <akiyks@gmail.com> wrot=
e:
> >>>
> > [...]
> >>> Yes, that should work.
> >>
> >> Ok, assigning to zero didn't work (it still complained about
> >> uninitialized read), but using a separate int *lenFail to assign to
> >> rLenPtr worked. Curiously, if I used rLenPtr =3D len1; in error case, =
it
> >> actually takes a bit more time to verify.
> >>
> >> So I've converted everything else as you suggested. I compiled latest
> >> herd7 and it doesn't produce any warnings. But it's also extremely
> >> slow, compared to the herd7 that I get by default. Validating simple
> >> 1p1c cases takes about 2.5x times longer (0.03s vs 0.07),
>
> Wait a moment!
>
> This 0.03s was the run time of the original 1p1c litmus test, wasn't it?
> Then you are comparing apples and oranges.
>
> How long does your default herd7 take to complete the updated 1p1c test?
>
>         Thanks, Akira

It could be new test vs old test, so I re-ran again. Identical
1p1c-unbound test:

OLD version:

$ herd7 -version && herd7 -unroll 0 -conf linux-kernel.cfg
../../Documentation/litmus-tests/bpf-rb/bpf-rb+1p1c+unbound.litmus
7.52, Rev: exported
Test bpf-rb+1p1c+unbound Allowed
States 2
0:rFail=3D0; 1:rFail=3D0; cx=3D0; len1=3D1; px=3D1;
0:rFail=3D0; 1:rFail=3D0; cx=3D1; len1=3D1; px=3D1;
Ok
Witnesses
Positive: 3 Negative: 0
Condition exists (0:rFail=3D0 /\ 1:rFail=3D0 /\ px=3D1 /\ len1=3D1 /\ (cx=
=3D0 \/ cx=3D1))
Observation bpf-rb+1p1c+unbound Always 3 0
Time bpf-rb+1p1c+unbound 0.03
Hash=3D20a68cc69b09fbb79f407f825c015623

LATEST from sources version:

$ herd7 -version && herd7 -unroll 0 -conf linux-kernel.cfg
../../Documentation/litmus-tests/bpf-rb/bpf-rb+1p1c+unbound.litmus
7.55+01(dev), Rev: 61e23aaee7bba87ccf4cdf1a620a3a9fa8f9a586
Test bpf-rb+1p1c+unbound Allowed
States 2
0:rFail=3D0; 1:rFail=3D0; cx=3D0; len1=3D1; px=3D1;
0:rFail=3D0; 1:rFail=3D0; cx=3D1; len1=3D1; px=3D1;
Ok
Witnesses
Positive: 3 Negative: 0
Condition exists (0:rFail=3D0 /\ 1:rFail=3D0 /\ px=3D1 /\ len1=3D1 /\ (cx=
=3D0 \/ cx=3D1))
Observation bpf-rb+1p1c+unbound Always 3 0
Time bpf-rb+1p1c+unbound 0.06
Hash=3D20a68cc69b09fbb79f407f825c015623

Still 2x difference.

>
> >>                                                           but trying
> >> to validate 2p1c case, which normally validates in 42s (unbounded) and
> >> 110s (bounded), it took more than 20 minutes and hasn't finished,
> >> before I gave up. So I don't know what's going on there...
> >
> > herdtools7 has recently been heavily restructured.
> > On the performance regression, I must defer to Luc.
> >
> > Luc, do you have any idea?
> >
> >>
> >> As for klitmus7, I managed to generate everything without warnings,
> >> but couldn't make it build completely due to:
> >>
> >> $ make
> >> make -C /lib/modules/5.6.13-01802-g938d64da97c6/build/
> >
> > So you are on Linux 5.6.x which requires cutting-edge klitmus7.
> >
> >> M=3D/home/andriin/local/linux-trees/tools/memory-model/mymodules modul=
es
> >> make[1]: Entering directory `/data/users/andriin/linux-build/fb-config=
'
> >> make[2]: Entering directory `/data/users/andriin/linux-build/default-x=
86_64'
> >>   CC [M]  /home/andriin/local/linux-trees/tools/memory-model/mymodules=
/litmus000.o
> >> /home/andriin/local/linux-trees/tools/memory-model/mymodules/litmus000=
.c:
> >> In function =E2=80=98zyva=E2=80=99:
> >> /home/andriin/local/linux-trees/tools/memory-model/mymodules/litmus000=
.c:507:12:
> >> warning: ISO C90 forbids variable length array =E2=80=98th=E2=80=99 [-=
Wvla]
> >>      struct task_struct *th[nth];
> >>             ^~~~~~~~~~~
> >> /home/andriin/local/linux-trees/tools/memory-model/mymodules/litmus000=
.c:
> >> In function =E2=80=98litmus_init=E2=80=99:
> >> /home/andriin/local/linux-trees/tools/memory-model/mymodules/litmus000=
.c:605:67:
> >> error: passing argument 4 of =E2=80=98proc_create=E2=80=99 from incomp=
atible pointer
> >> type [-Werror=3Dincompatible-pointer-types]
> >>    struct proc_dir_entry *litmus_pde =3D
> >> proc_create("litmus",0,NULL,&litmus_proc_fops);
> >>
> >> ^~~~~~~~~~~~~~~~~
> >> In file included from
> >> /home/andriin/local/linux-trees/tools/memory-model/mymodules/litmus000=
.c:15:
> >> /data/users/andriin/linux-fb/include/linux/proc_fs.h:64:24: note:
> >> expected =E2=80=98const struct proc_ops *=E2=80=99 but argument is of =
type =E2=80=98const
> >> struct file_operations *=E2=80=99
> >>  struct proc_dir_entry *proc_create(const char *name, umode_t mode,
> >> struct proc_dir_entry *parent, const struct proc_ops *proc_ops);
> >>                         ^~~~~~~~~~~
> >> cc1: some warnings being treated as errors
> >> make[3]: *** [/home/andriin/local/linux-trees/tools/memory-model/mymod=
ules/litmus000.o]
> >> Error 1
> >> make[2]: *** [/home/andriin/local/linux-trees/tools/memory-model/mymod=
ules]
> >> Error 2
> >> make[2]: Leaving directory `/data/users/andriin/linux-build/default-x8=
6_64'
> >> make[1]: *** [sub-make] Error 2
> >> make[1]: Leaving directory `/data/users/andriin/linux-build/fb-config'
> >> make: *** [all] Error 2
> >>
> >
> > These errors suggest the klitmus7 you used is version 7.52 or some such=
.
> > You said you have built herd7 from the source.  Have you also built kli=
tmus7?
> >
> > The up-to-date klitmus7 should generate code compatible with Linux 5.6.=
x.
> >
> > Could you try with the latest one?
> >
> >         Thanks, Akira
> >
> >>
> >> But at least it doesn't complain about atomic_t anymore. So anyways,
> >> I'm going to post updated litmus tests separately from BPF ringbuf
> >> patches, because Documentation/litmus-tests is not yet present in
> >> bpf-next.
> >>
> >>>
> >>> You can find a basic introduction of klitmus7 in tools/memory-model/R=
EADME.
> >>>
> >>>         Thanks, Akira
> >>>
> >>>>
> >>>>>
> >>>>> Please note that if you are on Linux 5.6 (or later), you need an up=
-to-date
> >>>>> klitmus7 due to a change in kernel API.
> >>>>>
> >>>>> Any question is welcome!
> >>>>>
> >>>>>         Thanks, Akira
> >>>>>
> >>
> >> [...]
> >>
> >
>