[PATCH bpf] bpf: powerpc64: optimize JIT passes for bpf function calls
From: Sandipan Das <hidden>
Date: 2018-12-03 12:24:21
Once the JITed images for each function in a multi-function program
are generated after the first three JIT passes, we only need to fix
the target address for the branch instruction corresponding to each
bpf-to-bpf function call.
This introduces the following optimizations for reducing the work
done by the JIT compiler when handling multi-function programs:
[1] Instead of doing two extra passes to fix the bpf function calls,
do just one as that would be sufficient.
[2] During the extra pass, only overwrite the instruction sequences
for the bpf-to-bpf function calls as everything else would still
remain exactly the same. This also reduces the number of writes
to the JITed image.
[3] Do not regenerate the prologue and the epilogue during the extra
pass as that would be redundant.
Signed-off-by: Sandipan Das <redacted>
---
arch/powerpc/net/bpf_jit_comp64.c | 66 +++++++++++++++++++++++++++++++
1 file changed, 66 insertions(+)
diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
index 17482f5de3e2..9393e231cbc2 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c@@ -891,6 +891,55 @@ static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, return 0; } +/* Fix the branch target addresses for subprog calls */ +static int bpf_jit_fixup_subprog_calls(struct bpf_prog *fp, u32 *image, + struct codegen_context *ctx, u32 *addrs) +{ + const struct bpf_insn *insn = fp->insnsi; + bool func_addr_fixed; + u64 func_addr; + u32 tmp_idx; + int i, ret; + + for (i = 0; i < fp->len; i++) { + /* + * During the extra pass, only the branch target addresses for + * the subprog calls need to be fixed. All other instructions + * can left untouched. + * + * The JITed image length does not change because we already + * ensure that the JITed instruction sequence for these calls + * are of fixed length by padding them with NOPs. + */ + if (insn[i].code == (BPF_JMP | BPF_CALL) && + insn[i].src_reg == BPF_PSEUDO_CALL) { + ret = bpf_jit_get_func_addr(fp, &insn[i], true, + &func_addr, + &func_addr_fixed); + if (ret < 0) + return ret; + + /* + * Save ctx->idx as this would currently point to the + * end of the JITed image and set it to the offset of + * the instruction sequence corresponding to the + * subprog call temporarily. + */ + tmp_idx = ctx->idx; + ctx->idx = addrs[i] / 4; + bpf_jit_emit_func_call_rel(image, ctx, func_addr); + + /* + * Restore ctx->idx here. This is safe as the length + * of the JITed sequence remains unchanged. + */ + ctx->idx = tmp_idx; + } + } + + return 0; +} + struct powerpc64_jit_data { struct bpf_binary_header *header; u32 *addrs;
@@ -989,6 +1038,22 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp) skip_init_ctx: code_base = (u32 *)(image + FUNCTION_DESCR_SIZE); + if (extra_pass) { + /* + * Do not touch the prologue and epilogue as they will remain + * unchanged. Only fix the branch target address for subprog + * calls in the body. + * + * This does not change the offsets and lengths of the subprog + * call instruction sequences and hence, the size of the JITed + * image as well. + */ + bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs); + + /* There is no need to perform the usual passes. */ + goto skip_codegen_passes; + } + /* Code generation passes 1-2 */ for (pass = 1; pass < 3; pass++) { /* Now build the prologue, body code & epilogue for real. */
@@ -1002,6 +1067,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp) proglen - (cgctx.idx * 4), cgctx.seen); } +skip_codegen_passes: if (bpf_jit_enable > 1) /* * Note that we output the base address of the code_base
--
2.17.2