Re: [PATCH v4 6/6] powerpc64/bpf: Additional NVR handling for bpf_throw

[Hari Bathini]

On 23/01/26 3:10 am, bot+bpf...@kernel.org wrote:
diff --git a/arch/powerpc/net/bpf_jit_comp64.c 
b/arch/powerpc/net/bpf_jit_comp64.c
index c284a12fce6e..98780a15fc69 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c

[ ... ]

@@ -125,9 +149,19 @@ int bpf_jit_stack_tailcallinfo_offset(struct 
codegen_context *ctx)

  static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg)
  {
-       if (reg >= BPF_PPC_NVR_MIN && reg < 32)
+       int min_valid_nvreg = BPF_PPC_NVR_MIN;
+       /* Default frame size for all cases except exception boundary */
+       int frame_nvr_size = BPF_PPC_STACKFRAME;
+
+       /* Consider all nv regs for handling exceptions */
+       if (ctx->exception_boundary || ctx->exception_cb) {
+               min_valid_nvreg = _R14;
+               frame_nvr_size = BPF_PPC_EXC_STACKFRAME;
+       }
+
+       if (reg >= min_valid_nvreg && reg < 32)
                return (bpf_has_stack_frame(ctx) ?
-                       (BPF_PPC_STACKFRAME + ctx->stack_size) : 0)
+                       (frame_nvr_size + ctx->stack_size) : 0)
                                - (8 * (32 - reg)) - BPF_PPC_TAILCALL;

For the redzone case (no stack frame) with exception_boundary, this
formula computes R23 at offset -80 and R24 at offset -72 from R1.

However, bpf_stf_barrier() uses hardcoded offsets -80(1) and -72(1) to
temporarily save R21 and R22 during STF barrier execution:

     asm (
     "     bpf_stf_barrier:                ;"
     "             std     21,-80(1)       ;"
     "             std     22,-72(1)       ;"
     "             sync                    ;"
     "             ld      21,-80(1)       ;"
     "             ld      22,-72(1)       ;"
     ...
     );

When an exception_boundary program without its own stack frame executes a
memory load that triggers STF_BARRIER_FALLBACK, can bpf_stf_barrier()
corrupt the saved R23 and R24 values? The hardcoded offsets would overlap
with the extended NVR save area.

If exception_cb later restores these registers, it would get R21/R22 values
instead of the original R23/R24 values.

That is practically not possible as exception_boundary would mean
at least bpf_throw() is in the program (SEEN_FUNC) and SEEN_FUNC
would setup a frame for the exception_boundary program. Also, 
bpf_stf_barrier() always uses the redzone. So, it never stomps
on the stack of exception_boundary program...

@abhishek, better add the above comment and probably
also make bpf_has_stack_frame() return true explicitly for
exception_boundary as well (though SEEN_FUNC can't be false
in case of exception_boundary)..

- Hari