bpf: Detect identical PTR_TO_MAP_VALUE_OR_NULL registers
authorThomas Graf <tgraf@suug.ch>
Tue, 18 Oct 2016 17:51:19 +0000 (19:51 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 22 Mar 2017 11:43:35 +0000 (12:43 +0100)
[ Upstream commit 57a09bf0a416700676e77102c28f9cfcb48267e0 ]

A BPF program is required to check the return register of a
map_elem_lookup() call before accessing memory. The verifier keeps
track of this by converting the type of the result register from
PTR_TO_MAP_VALUE_OR_NULL to PTR_TO_MAP_VALUE after a conditional
jump ensures safety. This check is currently exclusively performed
for the result register 0.

In the event the compiler reorders instructions, BPF_MOV64_REG
instructions may be moved before the conditional jump which causes
them to keep their type PTR_TO_MAP_VALUE_OR_NULL to which the
verifier objects when the register is accessed:

0: (b7) r1 = 10
1: (7b) *(u64 *)(r10 -8) = r1
2: (bf) r2 = r10
3: (07) r2 += -8
4: (18) r1 = 0x59c00000
6: (85) call 1
7: (bf) r4 = r0
8: (15) if r0 == 0x0 goto pc+1
 R0=map_value(ks=8,vs=8) R4=map_value_or_null(ks=8,vs=8) R10=fp
9: (7a) *(u64 *)(r4 +0) = 0
R4 invalid mem access 'map_value_or_null'

This commit extends the verifier to keep track of all identical
PTR_TO_MAP_VALUE_OR_NULL registers after a map_elem_lookup() by
assigning them an ID and then marking them all when the conditional
jump is observed.

Signed-off-by: Thomas Graf <tgraf@suug.ch>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
include/linux/bpf_verifier.h
kernel/bpf/verifier.c

index 6aaf425cebc39a9dbb3619d7ad28987ca54a01c7..7453c12815317575c96f271dedd7e76b2c7f2f10 100644 (file)
@@ -24,13 +24,13 @@ struct bpf_reg_state {
         */
        s64 min_value;
        u64 max_value;
+       u32 id;
        union {
                /* valid when type == CONST_IMM | PTR_TO_STACK | UNKNOWN_VALUE */
                s64 imm;
 
                /* valid when type == PTR_TO_PACKET* */
                struct {
-                       u32 id;
                        u16 off;
                        u16 range;
                };
index 8199821f54cf2c676204abb7b4cc9fad88524b4a..c428c9f85186fcfebefc98df34197b14d4b9e550 100644 (file)
@@ -212,9 +212,10 @@ static void print_verifier_state(struct bpf_verifier_state *state)
                else if (t == CONST_PTR_TO_MAP || t == PTR_TO_MAP_VALUE ||
                         t == PTR_TO_MAP_VALUE_OR_NULL ||
                         t == PTR_TO_MAP_VALUE_ADJ)
-                       verbose("(ks=%d,vs=%d)",
+                       verbose("(ks=%d,vs=%d,id=%u)",
                                reg->map_ptr->key_size,
-                               reg->map_ptr->value_size);
+                               reg->map_ptr->value_size,
+                               reg->id);
                if (reg->min_value != BPF_REGISTER_MIN_RANGE)
                        verbose(",min_value=%lld",
                                (long long)reg->min_value);
@@ -447,6 +448,7 @@ static void mark_reg_unknown_value(struct bpf_reg_state *regs, u32 regno)
 {
        BUG_ON(regno >= MAX_BPF_REG);
        regs[regno].type = UNKNOWN_VALUE;
+       regs[regno].id = 0;
        regs[regno].imm = 0;
 }
 
@@ -1252,6 +1254,7 @@ static int check_call(struct bpf_verifier_env *env, int func_id)
                        return -EINVAL;
                }
                regs[BPF_REG_0].map_ptr = meta.map_ptr;
+               regs[BPF_REG_0].id = ++env->id_gen;
        } else {
                verbose("unknown return type %d of func %d\n",
                        fn->ret_type, func_id);
@@ -1668,8 +1671,7 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
                                                insn->src_reg);
                                        return -EACCES;
                                }
-                               regs[insn->dst_reg].type = UNKNOWN_VALUE;
-                               regs[insn->dst_reg].map_ptr = NULL;
+                               mark_reg_unknown_value(regs, insn->dst_reg);
                        }
                } else {
                        /* case: R = imm
@@ -1931,6 +1933,38 @@ static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
        check_reg_overflow(true_reg);
 }
 
+static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
+                        enum bpf_reg_type type)
+{
+       struct bpf_reg_state *reg = &regs[regno];
+
+       if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
+               reg->type = type;
+               if (type == UNKNOWN_VALUE)
+                       mark_reg_unknown_value(regs, regno);
+       }
+}
+
+/* The logic is similar to find_good_pkt_pointers(), both could eventually
+ * be folded together at some point.
+ */
+static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
+                         enum bpf_reg_type type)
+{
+       struct bpf_reg_state *regs = state->regs;
+       int i;
+
+       for (i = 0; i < MAX_BPF_REG; i++)
+               mark_map_reg(regs, i, regs[regno].id, type);
+
+       for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
+               if (state->stack_slot_type[i] != STACK_SPILL)
+                       continue;
+               mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE,
+                            regs[regno].id, type);
+       }
+}
+
 static int check_cond_jmp_op(struct bpf_verifier_env *env,
                             struct bpf_insn *insn, int *insn_idx)
 {
@@ -2018,18 +2052,13 @@ static int check_cond_jmp_op(struct bpf_verifier_env *env,
        if (BPF_SRC(insn->code) == BPF_K &&
            insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
            dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
-               if (opcode == BPF_JEQ) {
-                       /* next fallthrough insn can access memory via
-                        * this register
-                        */
-                       regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
-                       /* branch targer cannot access it, since reg == 0 */
-                       mark_reg_unknown_value(other_branch->regs,
-                                              insn->dst_reg);
-               } else {
-                       other_branch->regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
-                       mark_reg_unknown_value(regs, insn->dst_reg);
-               }
+               /* Mark all identical map registers in each branch as either
+                * safe or unknown depending R == 0 or R != 0 conditional.
+                */
+               mark_map_regs(this_branch, insn->dst_reg,
+                             opcode == BPF_JEQ ? PTR_TO_MAP_VALUE : UNKNOWN_VALUE);
+               mark_map_regs(other_branch, insn->dst_reg,
+                             opcode == BPF_JEQ ? UNKNOWN_VALUE : PTR_TO_MAP_VALUE);
        } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
                   dst_reg->type == PTR_TO_PACKET &&
                   regs[insn->src_reg].type == PTR_TO_PACKET_END) {