bpf: improve verifier branch analysis
authorAlexei Starovoitov <ast@kernel.org>
Tue, 4 Dec 2018 06:46:05 +0000 (22:46 -0800)
committerDaniel Borkmann <daniel@iogearbox.net>
Tue, 4 Dec 2018 16:22:02 +0000 (17:22 +0100)
pathological bpf programs may try to force verifier to explode in
the number of branch states:
  20: (d5) if r1 s<= 0x24000028 goto pc+0
  21: (b5) if r0 <= 0xe1fa20 goto pc+2
  22: (d5) if r1 s<= 0x7e goto pc+0
  23: (b5) if r0 <= 0xe880e000 goto pc+0
  24: (c5) if r0 s< 0x2100ecf4 goto pc+0
  25: (d5) if r1 s<= 0xe880e000 goto pc+1
  26: (c5) if r0 s< 0xf4041810 goto pc+0
  27: (d5) if r1 s<= 0x1e007e goto pc+0
  28: (b5) if r0 <= 0xe86be000 goto pc+0
  29: (07) r0 += 16614
  30: (c5) if r0 s< 0x6d0020da goto pc+0
  31: (35) if r0 >= 0x2100ecf4 goto pc+0

Teach verifier to recognize always taken and always not taken branches.
This analysis is already done for == and != comparison.
Expand it to all other branches.

It also helps real bpf programs to be verified faster:
                       before  after
bpf_lb-DLB_L3.o         2003    1940
bpf_lb-DLB_L4.o         3173    3089
bpf_lb-DUNKNOWN.o       1080    1065
bpf_lxc-DDROP_ALL.o     29584   28052
bpf_lxc-DUNKNOWN.o      36916   35487
bpf_netdev.o            11188   10864
bpf_overlay.o           6679    6643
bpf_lcx_jit.o           39555   38437

Reported-by: Anatoly Trosinenko <anatoly.trosinenko@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
kernel/bpf/verifier.c
tools/testing/selftests/bpf/test_verifier.c

index 751bb30b7c5cc82b766110e930a3118868624dc8..55a49703f423e28689827575b7182bdcd06ba651 100644 (file)
@@ -3751,6 +3751,79 @@ static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
        }
 }
 
+/* compute branch direction of the expression "if (reg opcode val) goto target;"
+ * and return:
+ *  1 - branch will be taken and "goto target" will be executed
+ *  0 - branch will not be taken and fall-through to next insn
+ * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
+ */
+static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
+{
+       if (__is_pointer_value(false, reg))
+               return -1;
+
+       switch (opcode) {
+       case BPF_JEQ:
+               if (tnum_is_const(reg->var_off))
+                       return !!tnum_equals_const(reg->var_off, val);
+               break;
+       case BPF_JNE:
+               if (tnum_is_const(reg->var_off))
+                       return !tnum_equals_const(reg->var_off, val);
+               break;
+       case BPF_JGT:
+               if (reg->umin_value > val)
+                       return 1;
+               else if (reg->umax_value <= val)
+                       return 0;
+               break;
+       case BPF_JSGT:
+               if (reg->smin_value > (s64)val)
+                       return 1;
+               else if (reg->smax_value < (s64)val)
+                       return 0;
+               break;
+       case BPF_JLT:
+               if (reg->umax_value < val)
+                       return 1;
+               else if (reg->umin_value >= val)
+                       return 0;
+               break;
+       case BPF_JSLT:
+               if (reg->smax_value < (s64)val)
+                       return 1;
+               else if (reg->smin_value >= (s64)val)
+                       return 0;
+               break;
+       case BPF_JGE:
+               if (reg->umin_value >= val)
+                       return 1;
+               else if (reg->umax_value < val)
+                       return 0;
+               break;
+       case BPF_JSGE:
+               if (reg->smin_value >= (s64)val)
+                       return 1;
+               else if (reg->smax_value < (s64)val)
+                       return 0;
+               break;
+       case BPF_JLE:
+               if (reg->umax_value <= val)
+                       return 1;
+               else if (reg->umin_value > val)
+                       return 0;
+               break;
+       case BPF_JSLE:
+               if (reg->smax_value <= (s64)val)
+                       return 1;
+               else if (reg->smin_value > (s64)val)
+                       return 0;
+               break;
+       }
+
+       return -1;
+}
+
 /* Adjusts the register min/max values in the case that the dst_reg is the
  * variable register that we are working on, and src_reg is a constant or we're
  * simply doing a BPF_K check.
@@ -4152,21 +4225,15 @@ static int check_cond_jmp_op(struct bpf_verifier_env *env,
 
        dst_reg = &regs[insn->dst_reg];
 
-       /* detect if R == 0 where R was initialized to zero earlier */
-       if (BPF_SRC(insn->code) == BPF_K &&
-           (opcode == BPF_JEQ || opcode == BPF_JNE) &&
-           dst_reg->type == SCALAR_VALUE &&
-           tnum_is_const(dst_reg->var_off)) {
-               if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
-                   (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
-                       /* if (imm == imm) goto pc+off;
-                        * only follow the goto, ignore fall-through
-                        */
+       if (BPF_SRC(insn->code) == BPF_K) {
+               int pred = is_branch_taken(dst_reg, insn->imm, opcode);
+
+               if (pred == 1) {
+                        /* only follow the goto, ignore fall-through */
                        *insn_idx += insn->off;
                        return 0;
-               } else {
-                       /* if (imm != imm) goto pc+off;
-                        * only follow fall-through branch, since
+               } else if (pred == 0) {
+                       /* only follow fall-through branch, since
                         * that's where the program will go
                         */
                        return 0;
index 5dd4410a716ce1ac09466ed6c26c21b37b2cff84..df6f751cc1e81c0bcc1f68e86a5bc57e530a2452 100644 (file)
@@ -8576,7 +8576,7 @@ static struct bpf_test tests[] = {
                        BPF_JMP_IMM(BPF_JA, 0, 0, -7),
                },
                .fixup_map_hash_8b = { 4 },
-               .errstr = "R0 invalid mem access 'inv'",
+               .errstr = "unbounded min value",
                .result = REJECT,
        },
        {
@@ -10547,7 +10547,7 @@ static struct bpf_test tests[] = {
                "check deducing bounds from const, 5",
                .insns = {
                        BPF_MOV64_IMM(BPF_REG_0, 0),
-                       BPF_JMP_IMM(BPF_JSGE, BPF_REG_0, 0, 1),
+                       BPF_JMP_IMM(BPF_JSGE, BPF_REG_0, 1, 1),
                        BPF_ALU64_REG(BPF_SUB, BPF_REG_0, BPF_REG_1),
                        BPF_EXIT_INSN(),
                },