In order to generate the prologue and epilogue, the BPF JIT needs to
know which registers that are clobbered. Therefore, the during
pre-final passes, the prologue is generated after the body of the
program body-prologue-epilogue. Then, in the final pass, a proper
prologue-body-epilogue JITted image is generated.
This scheme has worked most of the time. However, for some large
programs with many jumps, e.g. the test_kmod.sh BPF selftest with
hardening enabled (blinding constants), this has shown to be
incorrect. For the final pass, when the proper prologue-body-epilogue
is generated, the image has not converged. This will lead to that the
final image will have incorrect jump offsets. The following is an
excerpt from an incorrect image:
| ...
| 3b8:
00c50663 beq a0,a2,3c4 <.text+0x3c4>
| 3bc:
0020e317 auipc t1,0x20e
| 3c0:
49630067 jalr zero,1174(t1) # 20e852 <.text+0x20e852>
| ...
| 20e84c: 8796 c.mv a5,t0
| 20e84e: 6422 c.ldsp s0,8(sp) # Epilogue start
| 20e850: 6141 c.addi16sp sp,16
| 20e852: 853e c.mv a0,a5 # Incorrect jump target
| 20e854: 8082 c.jr ra
The image has shrunk, and the epilogue offset is incorrect in the
final pass.
Correct the problem by always generating proper prologue-body-epilogue
outputs, which means that the first pass will only generate the body
to track what registers that are touched.
Fixes: 2353ecc6f91f ("bpf, riscv: add BPF JIT for RV64G")
Signed-off-by: Björn Töpel <bjorn@rivosinc.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20230710074131.19596-1-bjorn@kernel.org
struct bpf_prog *prog;
u16 *insns; /* RV insns */
int ninsns;
- int body_len;
+ int prologue_len;
int epilogue_offset;
int *offset; /* BPF to RV */
int nexentries;
int from, to;
off++; /* BPF branch is from PC+1, RV is from PC */
- from = (insn > 0) ? ctx->offset[insn - 1] : 0;
- to = (insn + off > 0) ? ctx->offset[insn + off - 1] : 0;
+ from = (insn > 0) ? ctx->offset[insn - 1] : ctx->prologue_len;
+ to = (insn + off > 0) ? ctx->offset[insn + off - 1] : ctx->prologue_len;
return ninsns_rvoff(to - from);
}
unsigned int prog_size = 0, extable_size = 0;
bool tmp_blinded = false, extra_pass = false;
struct bpf_prog *tmp, *orig_prog = prog;
- int pass = 0, prev_ninsns = 0, prologue_len, i;
+ int pass = 0, prev_ninsns = 0, i;
struct rv_jit_data *jit_data;
struct rv_jit_context *ctx;
prog = orig_prog;
goto out_offset;
}
+
+ if (build_body(ctx, extra_pass, NULL)) {
+ prog = orig_prog;
+ goto out_offset;
+ }
+
for (i = 0; i < prog->len; i++) {
prev_ninsns += 32;
ctx->offset[i] = prev_ninsns;
for (i = 0; i < NR_JIT_ITERATIONS; i++) {
pass++;
ctx->ninsns = 0;
+
+ bpf_jit_build_prologue(ctx);
+ ctx->prologue_len = ctx->ninsns;
+
if (build_body(ctx, extra_pass, ctx->offset)) {
prog = orig_prog;
goto out_offset;
}
- ctx->body_len = ctx->ninsns;
- bpf_jit_build_prologue(ctx);
+
ctx->epilogue_offset = ctx->ninsns;
bpf_jit_build_epilogue(ctx);
if (!prog->is_func || extra_pass) {
bpf_jit_binary_lock_ro(jit_data->header);
- prologue_len = ctx->epilogue_offset - ctx->body_len;
for (i = 0; i < prog->len; i++)
- ctx->offset[i] = ninsns_rvoff(prologue_len +
- ctx->offset[i]);
+ ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
bpf_prog_fill_jited_linfo(prog, ctx->offset);
out_offset:
kfree(ctx->offset);