Re: [PATCH bpf-next 6/7] libbpf: BPF Static Keys support

[Eduard Zingerman <eddyz87@gmail.com>]

How about a slightly different modification of the Anton's idea.
Suppose that, as before, there is a special map type:

    struct {
        __uint(type, BPF_MAP_TYPE_ARRAY);
        __type(key, __u32);
        __type(value, __u32);
        __uint(map_flags, BPF_F_STATIC_KEY);
        __uint(max_entries, 1);
    } skey1 SEC(".maps")

Which is used as below:

    __attribute__((naked))
    int foo(void) {
      asm volatile (
                    "r0 = %[skey1] ll;"
                    "if r0 != r0 goto 1f;"
                    "r1 = r10;"
                    "r1 += -8;"
                    "r2 = 1;"
                    "call %[bpf_trace_printk];"
            "1:"
                    "exit;"
                    :: __imm_addr(skey1),
                       __imm(bpf_trace_printk)
                    : __clobber_all
      );
    }

Disassembly of section .text:

0000000000000000 <foo>:
       0:   r0 = 0x0 ll
        0000000000000000:  R_BPF_64_64  skey1  ;; <---- Map relocation as usual
       2:   if r0 == r0 goto +0x4 <foo+0x38>   ;; <---- Note condition
       3:   r1 = r10
       4:   r1 += -0x8
       5:   r2 = 0x1
       6:   call 0x6
       7:   exit

And suppose that verifier is modified in the following ways:
- treat instructions "if rX == rX" / "if rX != rX" (when rX points to
  static key map) in a special way:
  - when program is verified, the jump is considered non deterministic;
  - when program is jitted, the jump is compiled as nop for "!=" and as
    unconditional jump for "==";
- build a table of static keys based on a specific map referenced in
  condition, e.g. for the example above it can be inferred that insn 2
  associates with map skey1 because "r0" points to "skey1";
- jit "rX = <static key> ll;" as nop;

On the plus side:
- any kinds of jump tables are omitted from system call;
- no new instruction is needed;
- almost no modifications to libbpf are necessary (only a helper macro
  to convince clang to keep "if rX == rX");

What do you think?

Thanks,
Eduard