Re: [PATCH bpf v2 2/6] bpf: powerpc64: add JIT support for multi-function programs

On 05/18/2018 06:05 PM, Naveen N. Rao wrote:

Daniel Borkmann wrote:

quoted

On 05/18/2018 02:50 PM, Sandipan Das wrote:

quoted

This adds support for bpf-to-bpf function calls in the powerpc64
JIT compiler. The JIT compiler converts the bpf call instructions
to native branch instructions. After a round of the usual passes,
the start addresses of the JITed images for the callee functions
are known. Finally, to fixup the branch target addresses, we need
to perform an extra pass.

Because of the address range in which JITed images are allocated
on powerpc64, the offsets of the start addresses of these images
from __bpf_call_base are as large as 64 bits. So, for a function
call, we cannot use the imm field of the instruction to determine
the callee's address. Instead, we use the alternative method of
getting it from the list of function addresses in the auxillary
data of the caller by using the off field as an index.

Signed-off-by: Sandipan Das <redacted>
---
 arch/powerpc/net/bpf_jit_comp64.c | 79 ++++++++++++++++++++++++++++++++++-----
 1 file changed, 69 insertions(+), 10 deletions(-)

diff --git a/arch/powerpc/net/bpf_jit_comp64.c b/arch/powerpc/net/bpf_jit_comp64.c
index 1bdb1aff0619..25939892d8f7 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c

@@ -256,7 +256,7 @@ static void bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32

 /* Assemble the body code between the prologue & epilogue */
 static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
                   struct codegen_context *ctx,
-                  u32 *addrs)
+                  u32 *addrs, bool extra_pass)
 {
     const struct bpf_insn *insn = fp->insnsi;
     int flen = fp->len;

@@ -712,11 +712,23 @@ static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,

             break;
 
         /*
-         * Call kernel helper
+         * Call kernel helper or bpf function
          */
         case BPF_JMP | BPF_CALL:
             ctx->seen |= SEEN_FUNC;
-            func = (u8 *) __bpf_call_base + imm;
+
+            /* bpf function call */
+            if (insn[i].src_reg == BPF_PSEUDO_CALL && extra_pass)

Perhaps it might make sense here for !extra_pass to set func to some dummy
address as otherwise the 'kernel helper call' branch used for this is a bit
misleading in that sense. The PPC_LI64() used in bpf_jit_emit_func_call()
optimizes the immediate addr, I presume the JIT can handle situations where
in the final extra_pass the image needs to grow/shrink again (due to different
final address for the call)?

That's a good catch. We don't handle that -- we expect to get the size right on first pass. We could probably have PPC_FUNC_ADDR() pad the result with nops to make it a constant 5-instruction sequence.

Yeah, arm64 does something similar by not optimizing the imm in order to always
emit 4 insns for it.

Thanks,
Daniel