Christophe Leroy wrote:

Le 04/10/2021 à 20:24, Naveen N. Rao a écrit :

quoted

Christophe Leroy wrote:

quoted

Le 01/10/2021 à 23:14, Naveen N. Rao a écrit :

quoted

In some scenarios, it is possible that the program epilogue is outside
the branch range for a BPF_EXIT instruction. Instead of rejecting such
programs, emit an indirect branch. We track the size of the bpf program
emitted after the initial run and do a second pass since BPF_EXIT can
end up emitting different number of instructions depending on the
program size.

Suggested-by: Jordan Niethe <redacted>
Signed-off-by: Naveen N. Rao <redacted>
---
  arch/powerpc/net/bpf_jit.h        |  3 +++
  arch/powerpc/net/bpf_jit_comp.c   | 22 +++++++++++++++++++++-
  arch/powerpc/net/bpf_jit_comp32.c |  2 +-
  arch/powerpc/net/bpf_jit_comp64.c |  2 +-
  4 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/arch/powerpc/net/bpf_jit.h b/arch/powerpc/net/bpf_jit.h
index 89bd744c2bffd4..4023de1698b9f5 100644
--- a/arch/powerpc/net/bpf_jit.h
+++ b/arch/powerpc/net/bpf_jit.h

@@ -126,6 +126,7 @@

  #define SEEN_FUNC    0x20000000 /* might call external helpers */
  #define SEEN_TAILCALL    0x40000000 /* uses tail calls */
+#define SEEN_BIG_PROG    0x80000000 /* large prog, >32MB */
  #define SEEN_VREG_MASK    0x1ff80000 /* Volatile registers r3-r12 */
  #define SEEN_NVREG_MASK    0x0003ffff /* Non volatile registers 
r14-r31 */

@@ -179,6 +180,8 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 

*image, struct codegen_context *
  void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx);
  void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx);
  void bpf_jit_realloc_regs(struct codegen_context *ctx);
+int bpf_jit_emit_exit_insn(u32 *image, struct codegen_context *ctx,
+                    int tmp_reg, unsigned long exit_addr);
  #endif

diff --git a/arch/powerpc/net/bpf_jit_comp.c 

b/arch/powerpc/net/bpf_jit_comp.c
index fcbf7a917c566e..3204872fbf2738 100644

--- a/arch/powerpc/net/bpf_jit_comp.c
+++ b/arch/powerpc/net/bpf_jit_comp.c

@@ -72,6 +72,21 @@ static int bpf_jit_fixup_subprog_calls(struct 

bpf_prog *fp, u32 *image,
      return 0;
  }
+int bpf_jit_emit_exit_insn(u32 *image, struct codegen_context *ctx,
+                    int tmp_reg, unsigned long exit_addr)
+{
+    if (!(ctx->seen & SEEN_BIG_PROG) && 
is_offset_in_branch_range(exit_addr)) {
+        PPC_JMP(exit_addr);
+    } else {
+        ctx->seen |= SEEN_BIG_PROG;
+        PPC_FUNC_ADDR(tmp_reg, (unsigned long)image + exit_addr);
+        EMIT(PPC_RAW_MTCTR(tmp_reg));
+        EMIT(PPC_RAW_BCTR());
+    }
+
+    return 0;
+}
+
  struct powerpc64_jit_data {
      struct bpf_binary_header *header;
      u32 *addrs;

@@ -155,12 +170,17 @@ struct bpf_prog *bpf_int_jit_compile(struct 

bpf_prog *fp)
          goto out_addrs;
      }
+    if (!is_offset_in_branch_range((long)cgctx.idx * 4))
+        cgctx.seen |= SEEN_BIG_PROG;
+
      /*
       * If we have seen a tail call, we need a second pass.
       * This is because bpf_jit_emit_common_epilogue() is called
       * from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
+     * We also need a second pass if we ended up with too large
+     * a program so as to fix branches.
       */
-    if (cgctx.seen & SEEN_TAILCALL) {
+    if (cgctx.seen & (SEEN_TAILCALL | SEEN_BIG_PROG)) {
          cgctx.idx = 0;
          if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
              fp = org_fp;

diff --git a/arch/powerpc/net/bpf_jit_comp32.c 

b/arch/powerpc/net/bpf_jit_comp32.c
index a74d52204f8da2..d2a67574a23066 100644

--- a/arch/powerpc/net/bpf_jit_comp32.c
+++ b/arch/powerpc/net/bpf_jit_comp32.c

@@ -852,7 +852,7 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 

*image, struct codegen_context *
               * we'll just fall through to the epilogue.
               */
              if (i != flen - 1)
-                PPC_JMP(exit_addr);
+                bpf_jit_emit_exit_insn(image, ctx, tmp_reg, exit_addr);

On ppc32, if you use tmp_reg you must flag it. But I think you could 
use r0 instead.

Indeed. Can we drop tracking of the temp registers and using them while
remapping registers? Are you seeing significant benefits with re-use of 
those temp registers?

I'm not sure to follow you.

On ppc32, all volatile registers are used for function arguments, so 
temp registers are necessarily non-volatile so we track them as all 
non-volatile registers we use.

I think saving on stack only the non-volatile registers we use provides 
real benefit, otherwise you wouldn't have implemented it would you ?

You're right. I was wary of having to track temporary register usage, 
which is a bit harder and prone to mistakes like the above. A related 
concern was that the register remapping is only used if there are no 
helper calls, which looks like a big limitation.

But, I do agree that it is worth the trouble for ppc32 given the 
register usage.

Anyway here you should use _R0 instead of tmp_reg.

Thanks,
Naveen