bpf: track spill/fill of constants
authorAlexei Starovoitov <ast@kernel.org>
Sat, 15 Jun 2019 19:12:17 +0000 (12:12 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Wed, 19 Jun 2019 00:22:51 +0000 (02:22 +0200)
commitf7cf25b2026dc8441e0fa3a202c2aa8a56211e30
treeae34479da9d5cad3485d0cd084e6a750cd600bf4
parenta324aae32fa9bfdd03e89078e20ebcbd7737fda5
bpf: track spill/fill of constants

Compilers often spill induction variables into the stack,
hence it is necessary for the verifier to track scalar values
of the registers through stack slots.

Also few bpf programs were incorrectly rejected in the past,
since the verifier was not able to track such constants while
they were used to compute offsets into packet headers.

Tracking constants through the stack significantly decreases
the chances of state pruning, since two different constants
are considered to be different by state equivalency.
End result that cilium tests suffer serious degradation in the number
of states processed and corresponding verification time increase.

                     before  after
bpf_lb-DLB_L3.o      1838    6441
bpf_lb-DLB_L4.o      3218    5908
bpf_lb-DUNKNOWN.o    1064    1064
bpf_lxc-DDROP_ALL.o  26935   93790
bpf_lxc-DUNKNOWN.o   34439   123886
bpf_netdev.o         9721    31413
bpf_overlay.o        6184    18561
bpf_lxc_jit.o        39389   359445

After further debugging turned out that cillium progs are
getting hurt by clang due to the same constant tracking issue.
Newer clang generates better code by spilling less to the stack.
Instead it keeps more constants in the registers which
hurts state pruning since the verifier already tracks constants
in the registers:
                  old clang  new clang
                         (no spill/fill tracking introduced by this patch)
bpf_lb-DLB_L3.o      1838    1923
bpf_lb-DLB_L4.o      3218    3077
bpf_lb-DUNKNOWN.o    1064    1062
bpf_lxc-DDROP_ALL.o  26935   166729
bpf_lxc-DUNKNOWN.o   34439   174607
bpf_netdev.o         9721    8407
bpf_overlay.o        6184    5420
bpf_lcx_jit.o        39389   39389

The final table is depressing:
                  old clang  old clang    new clang  new clang
                           const spill/fill        const spill/fill
bpf_lb-DLB_L3.o      1838    6441          1923      8128
bpf_lb-DLB_L4.o      3218    5908          3077      6707
bpf_lb-DUNKNOWN.o    1064    1064          1062      1062
bpf_lxc-DDROP_ALL.o  26935   93790         166729    380712
bpf_lxc-DUNKNOWN.o   34439   123886        174607    440652
bpf_netdev.o         9721    31413         8407      31904
bpf_overlay.o        6184    18561         5420      23569
bpf_lxc_jit.o        39389   359445        39389     359445

Tracking constants in the registers hurts state pruning already.
Adding tracking of constants through stack hurts pruning even more.
The later patch address this general constant tracking issue
with coarse/precise logic.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
kernel/bpf/verifier.c