According to the prototype formal BPF memory consistency model
discussed e.g. in [1] and following the ordering properties of
the C/in-kernel macro atomic_cmpxchg(), a BPF atomic operation
with the BPF_CMPXCHG modifier is fully ordered. However, the
current RISC-V JIT lowerings fail to meet such memory ordering
property. This is illustrated by the following litmus test:
BPF BPF__MP+success_cmpxchg+fence
{
0:r1=x; 0:r3=y; 0:r5=1;
1:r2=y; 1:r4=f; 1:r7=x;
}
P0 | P1 ;
*(u64 *)(r1 + 0) = 1 | r1 = *(u64 *)(r2 + 0) ;
r2 = cmpxchg_64 (r3 + 0, r4, r5) | r3 = atomic_fetch_add((u64 *)(r4 + 0), r5) ;
| r6 = *(u64 *)(r7 + 0) ;
exists (1:r1=1 /\ 1:r6=0)
whose "exists" clause is not satisfiable according to the BPF
memory model. Using the current RISC-V JIT lowerings, the test
can be mapped to the following RISC-V litmus test:
RISCV RISCV__MP+success_cmpxchg+fence
{
0:x1=x; 0:x3=y; 0:x5=1;
1:x2=y; 1:x4=f; 1:x7=x;
}
P0 | P1 ;
sd x5, 0(x1) | ld x1, 0(x2) ;
L00: | amoadd.d.aqrl x3, x5, 0(x4) ;
lr.d x2, 0(x3) | ld x6, 0(x7) ;
bne x2, x4, L01 | ;
sc.d x6, x5, 0(x3) | ;
bne x6, x4, L00 | ;
fence rw, rw | ;
L01: | ;
exists (1:x1=1 /\ 1:x6=0)
where the two stores in P0 can be reordered. Update the RISC-V
JIT lowerings/implementation of BPF_CMPXCHG to emit an SC with
RELEASE ("rl") annotation in order to meet the expected memory
ordering guarantees. The resulting RISC-V JIT lowerings of
BPF_CMPXCHG match the RISC-V lowerings of the C atomic_cmpxchg().
Other lowerings were fixed via 20a759df3b ("riscv, bpf: make
some atomic operations fully ordered").
Fixes: dd642ccb45 ("riscv, bpf: Implement more atomic operations for RV64")
Signed-off-by: Andrea Parri <parri.andrea@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Puranjay Mohan <puranjay@kernel.org>
Acked-by: Björn Töpel <bjorn@kernel.org>
Link: https://lpc.events/event/18/contributions/1949/attachments/1665/3441/bpfmemmodel.2024.09.19p.pdf [1]
Link: https://lore.kernel.org/bpf/20241017143628.2673894-1-parri.andrea@gmail.com
e59db0623f