Re: [PATCH bpf-next 1/1] bpf: Allow fall back to interpreter for programs with stack size <= 512

[KaFai Wan <kafai.wan@linux.dev>]

On Thu, 2025-08-07 at 09:50 -0700, Alexei Starovoitov wrote:
> On Tue, Aug 5, 2025 at 4:55 AM KaFai Wan <kafai.wan@linux.dev> wrote:
> > 
> > OpenWRT users reported regression on ARMv6 devices after updating
> > to latest
> > HEAD, where tcpdump filter:
> > 
> > tcpdump -i mon1 \
> > "not wlan addr3 3c37121a2b3c and not wlan addr2 184ecbca2a3a \
> > and not wlan addr2 14130b4d3f47 and not wlan addr2 f0f61cf440b7 \
> > and not wlan addr3 a84b4dedf471 and not wlan addr3 d022be17e1d7 \
> > and not wlan addr3 5c497967208b and not wlan addr2 706655784d5b"
> > 
> > fails with warning: "Kernel filter failed: No error information"
> > when using config:
> >  # CONFIG_BPF_JIT_ALWAYS_ON is not set
> >  CONFIG_BPF_JIT_DEFAULT_ON=y
> > 
> > The issue arises because commits:
> > 1. "bpf: Fix array bounds error with may_goto" changed default
> > runtime to
> >    __bpf_prog_ret0_warn when jit_requested = 1
> > 2. "bpf: Avoid __bpf_prog_ret0_warn when jit fails" returns error
> > when
> >    jit_requested = 1 but jit fails
> > 
> > This change restores interpreter fallback capability for BPF
> > programs with
> > stack size <= 512 bytes when jit fails.
> > 
> > Reported-by: Felix Fietkau <nbd@nbd.name>
> > Closes:
> > https://lore.kernel.org/bpf/2e267b4b-0540-45d8-9310-e127bf95fc63@nbd.name/
> > Fixes: 6ebc5030e0c5 ("bpf: Fix array bounds error with may_goto")
> 
> This commit looks fine.
> 
> > Fixes: 86bc9c742426 ("bpf: Avoid __bpf_prog_ret0_warn when jit
> > fails")
> 
> But this one is indeed problematic.
> But before we revert, please provide a selftest that is causing
> valid classic bpf prog to fail JITing on arm,
> because it has to be fixed as well. 
> 
OK, I'll add a test for it.

> Sounds like OpenWRT was suffering performance loss due to the
> interpreter.
> 
> > Signed-off-by: KaFai Wan <kafai.wan@linux.dev>
> > ---
> >  kernel/bpf/core.c | 12 +++++++-----
> >  1 file changed, 7 insertions(+), 5 deletions(-)
> > 
> > diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
> > index 5d1650af899d..2d86bd4b0b97 100644
> > --- a/kernel/bpf/core.c
> > +++ b/kernel/bpf/core.c
> > @@ -2366,8 +2366,8 @@ static unsigned int
> > __bpf_prog_ret0_warn(const void *ctx,
> >                                          const struct bpf_insn
> > *insn)
> >  {
> >         /* If this handler ever gets executed, then
> > BPF_JIT_ALWAYS_ON
> > -        * is not working properly, or interpreter is being used
> > when
> > -        * prog->jit_requested is not 0, so warn about it!
> > +        * or may_goto may cause stack size > 512 is not working
> > properly,
> > +        * so warn about it!
> 
> We shouldn't have touched this comment. Let's not do it again.
> 
OK.
> >          */
> >         WARN_ON_ONCE(1);
> >         return 0;
> > @@ -2478,10 +2478,10 @@ static void bpf_prog_select_func(struct
> > bpf_prog *fp)
> >          * But for non-JITed programs, we don't need bpf_func, so
> > no bounds
> >          * check needed.
> >          */
> > -       if (!fp->jit_requested &&
> > -           !WARN_ON_ONCE(idx >= ARRAY_SIZE(interpreters))) {
> > +       if (idx < ARRAY_SIZE(interpreters)) {
> >                 fp->bpf_func = interpreters[idx];
> 
> this is fine.
> 
> >         } else {
> > +               WARN_ON_ONCE(!fp->jit_requested);
> 
> drop it. Let's not give syzbot more opportunities
> to spam us again with fault injection -like corner cases.
> 
OK, will drop it.

> >                 fp->bpf_func = __bpf_prog_ret0_warn;
> >         }
> >  #else
> > @@ -2505,7 +2505,7 @@ struct bpf_prog
> > *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
> >         /* In case of BPF to BPF calls, verifier did all the prep
> >          * work with regards to JITing, etc.
> >          */
> > -       bool jit_needed = fp->jit_requested;
> > +       bool jit_needed = false;
> 
> ok
> 
> > 
> >         if (fp->bpf_func)
> >                 goto finalize;
> > @@ -2515,6 +2515,8 @@ struct bpf_prog
> > *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
> >                 jit_needed = true;
> > 
> >         bpf_prog_select_func(fp);
> > +       if (fp->bpf_func == __bpf_prog_ret0_warn)
> > +               jit_needed = true;
> 
> This is too hacky.
> Change bpf_prog_select_func() to return bool and
> rename it bpf_prog_select_func/bpf_prog_select_interpreter()
> 
> true on success, false on when interpreter is impossible.
> 
OK, will change it.

> And target bpf tree.
> 
OK. 
> --
> pw-bot: cr
> 
> > 
> >         /* eBPF JITs can rewrite the program in case constant
> >          * blinding is active. However, in case of error during
> > --
> > 2.43.0
> > 

-- 
Thanks,
KaFai